Cognitive functions of intracellular mechanisms for contextual amplification

Brain and Cognition xxx (2015) xxx–xxx

Contents lists available at ScienceDirect

Brain and Cognition journal homepage: www.elsevier.com/locate/b&c

Cognitive functions of intracellular mechanisms for contextual amplification William A. Phillips School of Natural Sciences, University of Stirling, Stirling FK9 4LA, UK

a r t i c l e

i n f o

Article history: Received 21 May 2015 Revised 16 September 2015 Accepted 18 September 2015 Available online xxxx Keywords: Apical amplification Disamplification Contextual modulation Consciousness Cognition Neocortical computation

a b s t r a c t Evidence for the hypothesis that input to the apical tufts of neocortical pyramidal cells plays a central role in cognition by amplifying their responses to feedforward input is reviewed. Apical tufts are electrically remote from the soma, and their inputs come from diverse sources including direct feedback from higher cortical regions, indirect feedback via the thalamus, and long-range lateral connections both within and between cortical regions. This suggests that input to tuft dendrites may amplify the cell’s response to basal inputs that they receive via layer 4 and which have synapses closer to the soma. ERP data supporting this inference is noted. Intracellular studies of apical amplification (AA) and of disamplification by inhibitory interneurons targeted only at tufts are reviewed. Cognitive processes that have been related to them by computational, electrophysiological, and psychopathological studies are then outlined. These processes include: figure-ground segregation and Gestalt grouping; contextual disambiguation in perception and sentence comprehension; priming; winner-take-all competition; attention and working memory; setting the level of consciousness; cognitive control; and learning. It is argued that theories in cognitive neuroscience should not assume that all neurons function as integrate-and-fire point processors, but should use the capabilities of cells with distinct sites of integration for driving and modulatory inputs. Potentially ‘unifying’ theories that depend upon these capabilities are reviewed. It is concluded that evolution of the primitives of AA and disamplification in neocortex may have extended cognitive capabilities beyond those built from the long-established primitives of excitation, inhibition, and disinhibition. Ó 2015 Elsevier Inc. All rights reserved.

1. Introduction Neocortical function depends on inference – inferences about distal things from proximal signals, inferences about the implications of symbolic propositions or beliefs, and inferences about the likely consequences of possible actions. As these inferences are often probabilistic and dependent upon the particular context in which they are made, neural systems must combine contextsensitivity with the dynamic coordination of many widely distributed local neuronal inferences, and ways in which this may be done are a topic of intense investigation (e.g. von der Malsburg, Phillips, & Singer, 2010). It is now well-established that this coordination involves various forms of contextual modulation (e.g. Lamme, 2004; Salinas & Sejnowski, 2001). Some amplify and group relevant activities; others suppress irrelevant activities. Sherman (2012) reviews physiological evidence showing that there are two clearly distinct classes of thalamocortical and corticocortical synaptic interaction; one is driving, the other is modulatory. E-mail address: [email protected]

Gilbert and Sigman (2007) review neurophysiological studies of top-down modulation in relation to sensory and higher processes. The functions, mechanisms, and malfunctions of contextual modulation in general have been reviewed recently (Phillips, Clark, & Silverstein, 2015), so this paper is focussed specifically on intracellular mechanisms that amplify response to driving inputs and on the cognitive functions that they are thought to have. It has often been argued that, though there are important variations, similarities in microcircuit anatomy and physiology across regions and species suggest that there may be a common (i.e. much used though not necessarily ubiquitous) neocortical strategy by which diverse types of information are processed, including those involved in sensation from diverse modalities, motor control and higher cognitive processes. For an assessment of this issue by many expert commentators see Phillips and Singer (1997). For a recent in-depth review supporting such arguments see Harris and Shepherd (2015). It has also been argued that these commonalities include regions classified as affective as well as those classified as cognitive (Pessoa, 2008), regions that are concerned with the theory of mind as well as those that perform basic sensorimotor

http://dx.doi.org/10.1016/j.bandc.2015.09.005 0278-2626/Ó 2015 Elsevier Inc. All rights reserved.

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

2

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

functions (Ondobaka, Kilner, & Friston, this issue), and regions concerned with phenomenological experiences such as the conscious sense of presence (Seth, 2013; Seth, Suzuki, & Critchley, 2011). So, how could a common information processing strategy serve such a wide variety of different purposes? As answers to this question cannot be framed using concepts specific to each of the domains of application they must be framed using more abstract concepts, such as those of information theory (Kay & Phillips, 2011; Wibral, Priesemann, Kay, Lizier, & Phillips, this issue) and probabilistic inference (Fiorillo, 2012; Friston, 2010; Phillips, 2012). Common abstract goals that are of most relevance to the intracellular mechanisms reviewed here can be described informally as amplifying and grouping signals that are relevant in the context of current activity elsewhere and as suppressing signals that are irrelevant. The concept of ‘grouping’ referred to here is that of dynamic grouping as defined by Phillips, von der Malsburg, and Singer (2010) which is a refined version of the notion of ‘binding’ and which is similar to the concept of ‘temporary coalitions’ (Crick & Koch, 2003). Section 4 outlines theories that express these goals as a formal objective for neural systems composed of local cortical processors that receive two functionally distinct sources of input: driving input from which the information to be transmitted is selected; and modulatory input that conveys contextual information that amplifies, suppresses, and coordinates responses to the drive. The central hypothesis examined here is that modulatory mechanisms within pyramidal neurons of the neocortex provide computational capabilities that this requires. This hypothesis contrasts with the common assumption that the neural bases of cognition can be adequately understood in terms of integrate-and-fire point processors. Such point processors sum their excitatory and inhibitory inputs and transmit a signal when that sum exceeds a threshold. Assuming that neurons function as point processors is useful in that it frees cognitive theorists from any need to grapple with the many complexities of intracellular processes. That assumption is misleading, however. The evidence reviewed here clearly indicates that perception, thought and action may be more correctly, and more economically, understood as involving pyramidal cells that, in addition to the somatic integration site for driving inputs, have a distinct apical integration site for modulatory inputs, with far-reaching consequences for cognition. The conception of ‘contextual modulation’ on which this paper is based was first defined rigorously by Phillips, Kay, and Smyth (1995). Further refinements and implications of that neuroncentric conception are discussed at length by Phillips et al. (2015, Section 2). The most recent development in this perspective on contextual modulation builds upon recent advances in information theory. Those advances extend information theory beyond the case of mutual information between input and output variables, and provide general definitions of unique, shared, and synergistic components of transmitted information for the case of multiple inputs. For a formal presentation of that advance and its relevance to contextual modulation see Wibral et al. (this issue). In essence, modulatory inputs are distinguished from driving inputs on the grounds that, whereas driving signals can produce an output by themselves, the effects of modulatory inputs are conditional upon the presence of a primary driving signal to which response is modulated. A key information-theoretic criterion for this modulation is that information transmitted uniquely about the modulator increases as the strength of another input variable (i.e. that to which response is modulated) is increased from zero (Phillips & Craven, 2000; Smyth, Phillips, & Kay, 1996). If the strength of the driving receptive field input is denoted by R, the strength of the modulatory contextual field input by C, and the output strength by Y then a simple form of contextual modulation is Y = R + RC. This contains the necessary asymmetry between R and C because R can then

produce an output when C = 0 but C cannot produce an output when R = 0. It also distinguishes this form of contextual modulation from a purely multiplicative interaction in which Y = RC. In that latter case the multiplicative interaction is not amplifying a response that signals the presence of R but is computing the value of a new variable that requires the presence of both R and C, as in coordinate transformation, for example (Phillips & Silverstein, 2013). See Section 4.3 for further discussion of this issue. As the focus of this paper is on modulatory interactions within the neocortex it is important to see how they are related to the well-known effects of the classical neuromodulators, such as norepinephrine, and dopamine. The fundamental difference between modulatory interactions within the neocortex, such as those implemented by AA, and modulation of cortical activity by subcortical neuromodulators is that, whereas sub-cortical modulation is diffuse, intrinsic modulatory interactions, such as AA, are far more specific, thus enabling the context-sensitive selection of particular percepts, thoughts, and actions. These intra-cortical modulatory interactions may be of especial importance to human cognition, which seems to excel in such context-sensitive flexibility. As our understanding of locally specific intra-cortical modulatory processes grows many questions will arise concerning their interaction with diffuse neuromodulation. The following sections are organized as follows. Section 2 summarizes direct and indirect evidence indicating that some classes of pyramidal cell have intracellular mechanisms for contextsensitive amplification, and that particular types of inhibitory interneuron have evolved to specifically regulate that amplification. Section 3 outlines various cognitive functions to which these mechanisms are thought to contribute. It also outlines a few of the many psychopathological disorders arising from their malfunction. Section 4 clarifies what is meant by ‘unifying’, and re-assesses some potentially unifying theories of cortical computation in the light of the evidence for AA. Section 5 lists some of the many unresolved issues that arise. 2. Intracellular mechanisms for contextual amplification Structure often provides a clue to function (Crick & Koch, 2003), so this section first seeks clues in pyramidal cell morphology and in aspects of long-range connectivity. Direct intracellular evidence supporting the inferences drawn from these clues is then reviewed. It suggests that input to synapses in layer 1 can amplify pyramidal cell responses, and that specific inhibitory mechanisms have evolved to regulate that amplification. 2.1. As inputs to the apical tuft are electrically remote from the soma they could have evolved modulatory capabilities Pyramidal cells of the neocortex have an apical trunk that ascends from the cell body, or soma, to a dendritic tree called the apical tuft. If their soma is in layers 2, 3, or 5 then the tuft is in layer 1 of the cortex. If the soma is in layer 6 then the tuft is in layer 4. Tuft dendrites are richly studded with synapses but they are electrically remote from the soma. Without active dendritic mechanisms the effects of synaptic input to the tuft would be so strongly attenuated at the soma that they would have little or no effect on action potential generation (Häusser & Mel, 2003). The apical trunk must therefore have mechanisms for active signal propagation. They may include mechanisms that compensate for the distance of the tuft in a way that enables it to contribute to a ‘dendritic democracy’ in which all synapses have an approximately equal opportunity to contribute to the cell’s output (Häusser & Mel, 2003; Magee & Cook, 2000; Williams & Stuart, 2003). Thus, the morphological asymmetry between basal and tuft dendrites does

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

not necessarily imply that there is also a functional asymmetry. It does provide an opportunity for it, however. The distance of tuft dendrites from the soma opens the possibility of a primary driving role for the basal dendrites and a modulatory role for the tuft. It may be that this opportunity is taken for some types of pyramidal cell in some conditions. There is evidence for variations in tuft function across species and stages of development. For example, in hippocampal neurons the average somatic amplitude of EPSPs is independent of the apical dendritic site of generation, but in layer 5 neocortical pyramidal cells the average somatic amplitude of excitatory post-synaptic potentials (EPSPs) reduces with distance in synapse location along the apical dendritic tree (Williams & Stuart, 2003). This suggests that apical input contributes to a dendritic democracy in hippocampal neurons but not in neocortical cells. Thus the functional role of tuft dendrites may change during the course of evolutionary history. In the more ancient 3-layer cortex of fish, amphibia, and reptiles and of mammalian hippocampus and olfactory cortex the apical dendrites are the main input site of primary feedforward drive, and lack many of the active properties found in neocortical neurons (Bathellier, Margrie, & Larkum, 2009). This does not imply that there are no non-linear dendritic interactions in 3-layer cortex because there are many forms of non-linear interaction other than those involved in AA. It is possible that some aspects of receptive field selectivity in both 3- and 6-layer cortex are specified by non-linear computations that are clearly distinct from those of AA. It is in six-layer mammalian neocortex that the modulatory capabilities of AA have been observed within pyramidal cells (Shepherd, 2011). Thus, novel capabilities for active dendritic propagation in the apical trunk may have evolved in neocortex, along with specialized mechanisms for disinhibiting them. 2.2. Apical tufts in layer 1 of neocortex receive inputs from diverse sources of contextual information The sources of their input provide another anatomical clue to the function of tuft dendrites. Basal dendrites of layer 3 pyramidal cells receive the cell’s feedforward input via layer 4 from the few compact sources that specify the cell’s selective sensitivity (Markov et al., 2014). In contrast to that the diverse inputs to pyramidal cell tufts include direct feedback from higher cortical regions, indirect feedback via the thalamus, and long-range lateral connections both within and between cortical regions (Gur & Snodderly, 2008; Rubio-Garrido, Perez-de-Manzo, Porrero, Galazo, & Clasca, 2009). About 90% of the synaptic inputs to layer 1 come from long-range connections, with only about 10% coming from nearby neurons. Though some layer 1 axons do connect to inhibitory interneurons, their synapses are predominantly excitatory (Shao & Burkhalter, 1996) and on the apical tufts of pyramidal cells (Budd, 1998). Studies of the exact locations of pyramidal cell input show that there is a high degree of differentiation such that feedforward input impinges on lower parts of the dendritic arborisation, whereas feedback from higher levels impinges on the upper parts (Petreanu, Mao, Sternson, & Svoboda, 2009). This all suggests that tuft synapses in layer 1 are likely to convey signals that are used for modulation. Modulation, as conceived here, is not identified with feedback, however. Although much feedback is modulatory, some may be driving. It is possible that feedback is driving relative to layer 2 pyramidal cells but modulatory relative to the layer 3 and layer 5 cells that provide the feedforward inputs to higher levels in the neocortical hierarchy (Markov et al., 2014). Thus, modulation should not be indentified with feedback (Markov & Kennedy, 2013). Furthermore, long-range lateral connections may be either driving or modulatory. Modulatory inputs are therefore defined by their effects, not by their source (Phillips et al., 2015).

3

Event related potential (ERP) data has also been interpreted as indicating that input to layer 1 provides modulatory contextual input. Scalp ERPs reflect both dendritic currents and action potentials, and there is a close correspondence between the intracortical ERP and the ERP at the overlying brain surface in sensory cortex (Shah et al., 2004). The occipital N1 component of the ERP has been associated with current sinks in cortical layers 1 and 2 and action potentials as deep as layer 5, and is thought to depend upon modulatory signals from higher regions (Cauller, 1995). This hypothesis is supported by evidence that N1 reflects modulation of sensory discrimination by executive control or other feedback processes (e.g. Fedota, McDonald, Roberts, & Parasuraman, 2012; Mehta, Ulbert, & Schroeder, 2000). Thus, these findings all suggest that pyramidal cell activity is modulated by distal apical inputs. 2.3. Intracellular evidence for apical amplification There is direct intracellular evidence on the effects of distal apical input. The threshold for current injection to initiate action potentials in layer 5 cells increases greatly with distance from the soma (Larkum, Senn, & Lüscher, 2004). This shows that apical input is greatly disadvantaged relative to basal inputs. Apical input that cannot alone initiate action potential output can greatly increase the cell’s response to its basal inputs (Larkum et al., 2004), however, so amplifying response to basal input may be its primary role. In vitro studies of rat somatosensory cortex show that, in addition to the somatic integration site that generates action potentials, pyramidal cells of layer 5 have an integration site at the top of the apical dendrite. If it receives adequate depolarization from the tuft dendrites and a backpropagated spike from the soma it triggers calcium spikes that travel down the apical trunk to the soma (Larkum, Nevian, Sandler, Polsky, & Schiller, 2009; Larkum, Zhu, & Sakmann, 1999). This backpropagation-activated calcium spike firing (BAC firing) has an explosive effect on the cell’s output because it amplifies response by turning a single action potential into a brief burst of 2–4 action potentials (M. Larkum, 2013). The output of several spikes within 20 ms may therefore signal a successful match between feedforward input to the somatic integration site and distal contextual input to layer 1. Intracellular amplification is reduced if NMDA receptors for the excitatory neurotransmitter glutamate are blocked (e.g. M. Larkum, 2013; Larkum et al., 2009; Palmer et al., 2014), so this is in harmony with evidence showing that NMDARs have a central role in contextual modulation (e.g. Phillips & Silverstein, 2003; Self, Kooijmansa, Supèr, Lamme, & Roelfsema, 2012). As this apical amplification could have far-reaching implications for our understanding of brain and cognition (Jadi, Behabadi, Poleg-Polsky, Schiller, & Mel, 2014; M. Larkum, 2013; Major, Larkum, & Schiller, 2013; Phillips et al., 2015), its essentials are outlined in Figs. 1 and 2. Apical amplification depends strongly upon the relative timing of basal and tuft depolarization. The time-window for maximal amplification is wider and earlier in cells of layer 5 than in those of layers 2 and 3 (Ledergerber & Larkum, 2012). Responses are only amplified if proximal depolarization occurs within about 10 msec of an impulsive tuft depolarization, which decays rapidly if not renewed. In somatosensory cortex of anesthetized rodents a brief electrical pulse to layer 1 amplifies the responses of layer 2, 3 and 5 pyramidal cells to their feedforward input if it occurs within the following 10 msec, but if it occurs 20–40 msec later then they are suppressed (Shlosberg, Amitai, & Azouz, 2006). Shlosberg et al. found that activation of layer 1 increased the magnitude and reliability of cortical LFPs, indicating that it increased synchronization of the local population of cells in addition to amplifying responses of the individual cells. This makes it more likely that the amplification will be observable in scalp potentials. The matching of

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

4

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

Diverse apical inputs

ʃ

a

modulatory interactions

Basal inputs

ʃ

AP

AP input Basal Apical low low low

mod.

mod.

low

mod.

mod. ||

AP output |

|

| |

|

|

|

| | | | | |

| |

| |

||| || || |||

||

s |||

output

Fig. 1. An idealized sketch of two functionally distinct sites of integration in pyramidal cells. ʃs is the somatic integration site that is necessary for the generation of action potentials (AP). It integrates depolarizing current received from inputs to basal and perisomatic synapses. Those inputs arise from sources that specify the cell’s selective sensitivity. ʃa is the apical integration site. Depolarization at this site is not necessary for AP generation but, given adequate depolarization at the somatic site, it can convert a single AP into a burst containing 2–4 APs and lasting about 10–20 msec (M. Larkum, 2013). Inputs to the apical integration site arise from a diverse variety of sources. An impression of the possible effects of these two sites on output is given on the right. AP input from basal and apical tuft synapses to each of the two sites is categorized into two levels: low, e.g. spontaneous firing rates; and moderate (mod.), e.g. mean rates seen under normal waking conditions. If basal input is low then a moderate level of apical input is shown as having little effect on output. If basal input is moderate, however, then a moderate level of apical input converts single APs into a burst. It may also help synchronize bursting across subsets of cells conveying population codes (not shown).

Level

Higher levels

ʃŜ

ʃ

a

ʃŜ

ʃ

a

N+1

ʃŜ

ʃ

ʃŜ

ʃ

a

N

ʃŜ

ʃ

N-1

a ?

ʃŜ

ʃ

a

Non-s Thalamus Amygdala etc

a

Fig. 2. Long-range sources of input to apical integration sites (ʃa) which are shown as ovals. These sources are shown for only one of the six local microcircuits though they are assumed to be essentially the same for each of the many millions of microcircuits of which human neocortex is composed. The combined sequence of feedforward computations within a local microcircuit is shown as ʃSˆ in two microcircuits at each of three different hierarchical levels. Outputs from the microcircuits are shown by thinner arrows than inputs to indicate that they compress information selected from their driving inputs. The apical integration sites of the microcircuits are shown as ovals, and inputs to their synapses are shown as arising from diverse sources (see review by Gur & Snodderly, 2008). They guide the selection process via both their effect on on-going processing and by their effect on learning. The figure shows two streams of processing that have little or no overlap between the sources of their feedforward input. These streams may be in either the same or different cortical regions. Representing the nodes in this way forces the location and function of each node’s inputs to be made explicit. This contrasts with the widely used practice of representing networks as connections between nodes assumed to be point processors. Note that, although both sites receive input, all output is via the somatic site.

proximal and distal inputs is often described as coincidence detection (e.g. M. Larkum, 2013) but that does not imply any asymmetry between the two inputs. It must therefore be emphasized that the coincidence is between inputs with different functional roles when apical input is modulatory, because that is fundamentally asymmetric (Phillips, 2012). BAC firing is stronger in cells of layer 5 than of other layers, but AA has also been observed in cells of layers 2 and 3 in rat somatosensory cortex (e.g. Larkum, Waters, Sakmann, & Helmchen, 2007; Ledergerber & Larkum, 2012; Palmer et al., 2014; Waters, Larkum, Sakmann, & Helmchen, 2003) and in prefrontal cortex (PFC) (Boudewijns et al., 2013). Though BAC firing seems to be the main mechanism by which AA is achieved in layer 5 cells, other mechanisms may be more prominent elsewhere. Pyramidal cells in layers 2 and 3 have both similarities and differences with those in layer 5 (Larkum et al., 2007). The main similarity is the existence of an integration site near the top of the apical dendrite that operates as a modulator. At least four differences may have important functional implications. First, their apical dendrites are substantially shorter and thinner than those of layer 5 cells (Ledergerber & Larkum, 2010). Second, in cells of layers 2 and 3 NMDA spikes in tuft dendrites can influence action potential generation in the absence of a back-propagated action potential (Palmer et al., 2014). Third, the apical dendrites of cells in layers 2 and 3 support only brief dendritic spikes. Consequently, they produce only a single additional action potential, not three or more. This lower ability of neurons in layers 2 and 3 to generate bursts of more than two action potentials is seen both in vitro and in vivo (De Kock, Bruno, Spors, & Sakmann, 2007). Finally, another difference is that layer 2 and 3 cells seem to lack the hyperpolarisation-activated currents that are common in the apical dendrites of layer 5 cells (Larkum et al., 2007). It may be that these differences reduce the need for amplification to be implemented by BAC firing in cells of layers 2 and 3. It is difficult to directly observe AA in vivo, but it has been done, e.g. in the motor cortex of anesthetized mice (Hill, Varga, Jia, Sakmann, & Konnerth, 2013). It is especially difficult to make relevant observations when animals are awake, but that has also been done, e.g. in layer 5 cells of rat somatosensory cortex (Murayama & Larkum, 2009), in rat prefrontal cortex (PFC) (Boudewijns et al., 2013), and in the barrel cortex of mice performing an objectlocalization task (Xu et al., 2012). Palmer et al. (2014) report that in awake mice NMDA spikes occur in the tuft dendrites of cells in layers 2 and 3 of somatosensory cortex, and enhance response to sensory input. Layer 1 stimulation alone did not evoke a Ca2+ response in tuft dendrites; but when paired with hindpaw stimulation a Ca2+ response was evoked on about 50% of trials, which was significantly more frequent than hindpaw stimulation alone. In accord with the in vitro findings, isolated NMDA spikes in single tuft branches had little influence on the cell’s output. The findings of Palmer et al clearly indicated that the NMDA-dependent transients producing these effects were not dependent on backpropagating action potentials. Thus, although the effectiveness of tuft depolarization in contributing to the cell’s output was still conditional upon the presence of feedforward input, AA in this case occurred without BAC firing (Palmer et al., 2014). 2.4. Inhibitory regulation of apical amplification Recurrent connections between excitatory cells require inhibitory regulation to prevent runaway over-activation. Different inhibitory interneurons target different parts of the pyramidal cells to which they are connected and various sub-types of inhibitory interneuron have evolved to regulate AA specifically (Gentet, 2012). This is crucial because the functional consequences of inhibiting (or disinhibiting) AA are fundamentally different from

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

inhibiting (or disinhibiting) the output of the cell as a whole. This section therefore briefly summarizes some of the key messages that seem to be emerging concerning the inhibitory regulation of AA. Gentet (2012) reviews eight subtypes of inhibitory interneuron in rodent barrel cortex. Three of these are particularly relevant here because they have dense axonal terminals in layer 1 (Wang et al., 2004). Thus these interneurons specifically target apical amplification without inhibiting the generation of action potentials (Larkum et al., 1999; Pérez-Garci, Gassmann, Bettler, & Larkum, 2006). It is therefore useful to distinguish the inhibition of apical amplification, or disamplification, from inhibition in general. In what has been called a tour de force of scientific enquiry (M.E. Larkum, 2013), Jiang, Wang, Lee, Stornetta, and Zhu (2013) showed that one class of layer 1 interneuron specifically inhibits amplification and another disinhibits it. Furthermore, Jiang et al tested these effects in the fully functioning in vivo network, and showed them to powerfully enhance or suppress calcium spiking in the apical dendrites of layer 5 pyramidal neurons. More specifically, one class of layer 1 interneuron (referred to as elongated neurogliaform cells) implements a form of lateral inhibition by suppressing amplification over a region large enough to contain several local columns, whereas the other (referred to as single bouquet cells) disinhibits layer 2, 3 and 5 pyramidal cells in a single local column (Jiang et al., 2013; Lee et al., 2014). Together these two classes of interneuron effectively focus amplification on the cells of all layers in single, sparsely distributed, columns. AA is also suppressed by Martinotti interneurons. Their soma and wide-spreading dendrites are in layers 2–6 and their extensive axonal arborisation is in layer 1 (Wang et al., 2004; Gentet, 2012). Despite their low numbers in a typical column their dense and widespread connectivity places them in an ideal position to control the general level of tuft excitability within the spread of their axonal arbour. Upper-layer Martinotti interneurons effectively suppress AA during quiet wakefulness, but can release it rapidly when needed. Lower-layer Martinotti neurons are activated by bursts in layer 5 pyramidal neurons and provide feedback inhibition of the apical integration site (Murayama et al., 2009). Inhibition targeted at the trunk of the apical dendrite can also produce disamplification. The reduction of sensitivity to one side of the body caused by concurrent stimulation to the equivalent location on the other side of the body is well-known to sensory physiologists and to neuropsychologists who study neglect. Palmer et al. (2012) studied the cellular and synaptic mechanisms involved in such interhemispheric inhibition by stimulating either one or both hind paws of rats while recording cellular activity in somatosensory cortex. They concluded that this form of interhemispheric inhibition is mediated predominantly through postsynaptic mechanisms in the apical trunks of pyramidal neurons. In the presence of interhemispheric inhibition a feedforward pulse produced only single spikes, whereas a burst was produced in its absence. Interhemispheric input had no effect on the cell’s output in the absence of feedforward input, thus contrasting with the effect of feedforward input in the absence of interhemispheric input. This clearly shows the asymmetry that is central to the distinction between driving and modulatory interactions. It is useful to distinguish disamplification from other forms of inhibition because there are interneurons that specifically target tufts. There seem to be few, if any, mechanisms that specifically disinhibit apical amplification, however. Unless interneurons are discovered that specifically inhibit disamplification, there will be no need for any concept of dis-disamplification (which would be a very indirect way of amplifying activity). 3. Putative cognitive functions of apical amplification The range of cognitive functions that have been suggested for AA is wide, but if it is common to pyramidal cells in many different

5

cortical regions then the range should be wide, perhaps even wider than outlined here. Given so wide a range it is not possible to review each possible function in depth, so references are given to relevant in-depth studies. This includes computer models that explicitly use AA (e.g. Brosch & Neumann, 2014; Körding & König, 2000; Siegel, Körding, & König, 2000; Spratling, 2002; Spratling & Johnson, 2004), or that rely upon the amplification that AA could provide (e.g. Kay, Floreano, & Phillips, 1998; Ullman, 1995). Each of the cognitive functions considered has been so intensively studied by perceptual and cognitive psychologists that they are often treated as distinct fields of research. These functions overlap, however, and they may well share common underlying goals and mechanisms. A crucial point to be remembered throughout is that AA is proposed to be a possible mechanism for each of these functions, not the one and only mechanism. 3.1. Figure-ground segregation and dynamic grouping The basic perceptual capabilities of figure-ground segregation and dynamic grouping have been much studied using various methodologies. Neurophysiological evidence clearly indicates that figure-ground segregation involves modulatory feedback that amplifies pyramidal cell responses to stimulus elements within or near the figural boundary (e.g. Zipser, Lamme, & Schiller, 1996). There has also long been evidence that contextual modulation includes processes that group elements into coherent wholes (Phillips & Silverstein, 2013; Phillips & Singer, 1997). Recent evidence provides further support for the hypothesis that contextual amplification of relevant signals plays a major role in these perceptual processes, and it has been argued that direct interactions between pyramidal cells via their apical dendrites is a simple way to implement this amplification (Häusser & Mel, 2003; M. Larkum, 2013; Phillips et al., 2015). For a recent neurocomputational model of dynamic grouping and other basic functions that explicitly uses the mechanism of AA see Brosch and Neumann (2014). 3.2. Contextual disambiguation It has often been hypothesized that contextual disambiguation is crucial to the perception of objects and words because it amplifies those interpretations of ambiguous parts that form a coherent whole (e.g. Phillips & Singer, 1997). In addition to the neurobiological evidence for it, this hypothesis is supported by computational studies. They have shown, for example, that interpretations of ambiguous data, such as the local junctions in line drawings of objects, can be guided by choosing local interpretations that provide a coherent description of the object as a whole (e.g. Clowes, 1971). Many psychological and neuroimaging studies provide further support. For example, psychophysical studies show that visual object recognition is greatly facilitated when objects are presented in the context of scenes in which they are likely to occur (Kveraga, Ghuman, & Bar, 2007). Studies that Kveraga et al. review include neuroimaging data indicating that a major multimodal association region, i.e. orbitofrontal cortex (OFC), provides some of the contextual information that is used to guide visual processing. Specific streams of processing convey low spatial frequency information rapidly from primary visual cortex to OFC. There it is combined with other information to compute possible interpretations of the objects being seen. These possible interpretations are then projected down to the object recognition regions in inferotemporal cortex (IT) where they are used to guide interpretation of both high and low spatial-frequency information arriving more slowly from the secondary visual areas. The initial guesses produced by OFC facilitate recognition by sensitizing IT to the most likely candidate

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

6

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

objects, thus enhancing the speed and accuracy of recognition. Of most importance here is the evidence that object recognition in IT is facilitated by amplifying those interpretations that are more likely given the context (Kveraga et al., 2007). Direct modulatory interactions between pyramidal cells via their apical dendrites provide an obvious mechanism by which this amplification could be achieved as demonstrated explicitly by the computational model of Siegel et al. (2000), and implicitly by those of Ullman (1995) and Kay et al. (1998). If it is indeed the case that AA is used to facilitate contextual disambiguation in visual object recognition then this may simply be one example of a far wider role for AA in contextual disambiguation. For example, amplifying modulation may also play a role in disambiguating visual motion signals, as shown in detail by the neurally inspired computational model of Bayerl and Neumann (2004). It may also have a role in language comprehension and sentence integration (Kveraga et al., 2007) where local ambiguity and its resolution using context are common (DeLong, Urbach, & Kutas, 2005; Duffy, 1986). To give just one example from the many available, Kotchoubey and El-Khoury (2014) show that the effects of sentence context can override the effects of large differences in the probability with which default meanings are ascribed to ambiguous words. Sub-dominant meanings that fit the context are automatically selected, while dominant meanings that do not are suppressed. As this requires mechanisms by which contextual inferences or predictions amplify the relevant and suppress the irrelevant they clearly show a need for capabilities that AA and disamplification provide. Effective and efficient mechanisms for contextual disambiguation are likely to be an important prerequisite for the evolution of linguistic capabilities because language is particularly replete with local ambiguity that is resolved holistically. It remains to be determined whether this involves the evolution of advanced forms of AA. 3.3. Priming Priming is an implicit memory effect in which exposure to one input influences response to another. Its influences are widespread throughout many cognitive domains, though they have been most studied in the visual domain (Wiggs & Martin, 1998). They overlap with the disambiguating processes just discussed, and can either facilitate or impede processing. Demonstrations of contextual priming show that recognizing one object facilitates recognition of other objects (e.g. Davenport & Potter, 2004; Palmer, 1975). Improved performance resulting from repeated presentation of an item is known as repetition priming and may involve enhanced synchrony between regions in the PFC and inferotemporal cortex (Ghuman, Bar, Dobbins, & Schnyer, 2008). As priming requires mechanisms for contextual modulation, intracellular mechanisms are an obvious candidate now that there is direct empirical evidence for them. AA is most obviously of potential relevance to facilitatory priming, but it could also be involved in negative priming, because that could depend upon mechanisms for disamplification. Different spatial distributions of these amplifying and disamplifying effects may provide one means whereby their possible contributions to priming could be tested. Functional neuroimaging during priming indicates that it is associated with a decrease in the neural activity elicited by primed items. This has been interpreted as implying that priming often leads to sparser, more selective, neural activity that is focused on task-relevant information (Wiggs & Martin, 1998), which clearly echoes the work reviewed in Section 2.4 showing that the inhibitory regulation of AA focuses it on single, sparsely distributed, columns (Jiang et al., 2013; Lee et al., 2014). A computational model of two-layer neural networks built from spiking neurons with the two intracellular sites of integration discovered by Larkum et al. (1999) shows explicitly how this could

produce effects analogous to priming (Siegel et al., 2000). Furthermore, an effective computational algorithm for object recognition based upon priming, relies upon amplifying interactions between ascending and descending signals of exactly the kind that AA could provide (Ullman, 1995). Intracellular mechanisms for contextual amplification are not necessary for priming, however, because models using microcircuits built from conventional Hodgkin–Huxley neurons with only a single site of integration can also simulate many priming phenomena (e.g. Moldakarimov, Bazhenov, & Sejnowski, 2010). Thus, as priming effects probably involve both intracellular and microcircuit mechanisms, the role of each in producing the various effects of priming remains to be determined. 3.4. Reciprocal inhibition between alternatives Reciprocal inhibition between alternatives is a computational theme repeated across many different cortical regions. One way in which it might be implemented is via suppression of apical amplification, e.g. via the elongated neurogliaform interneurons in layer 1 as outlined in Section 2.4. Evidence for such reciprocal inhibition in human cortex was reported by Allison, Puce, and McCarthy (2002), who studied event-related potential (ERP) recordings made directly from the surface of the fusiform gyrus and adjacent cortex in patients undergoing surgery for intractable epilepsy. They found that faces generated a positive ERP (P200) at sites that generated a word-specific negative ERP at about 200 msec (N200). Conversely, words generated a P200 at sites that had a face specific N200. To interpret these results they noted that there is evidence that N200 reflects excitatory depolarizing potentials in the apical dendrites of pyramidal cells (Mehta et al., 2000). They then argued that P200 ERPs reflect hyperpolarizing inhibition of apical dendrites, which implies that there is competition between cells with different selective sensitivities. They concluded that lateral inhibition of each others apical depolarizing currents would be a simple way to implement this competitive interaction, and would produce the ERPs observed. The main functions proposed for this mechanism were to sharpen category selectivity and to force a choice between alternative categorizations. Such a mechanism would differ from that in which reciprocal inhibition operates on the somatic compartment because it would leave feedforward transmission operating, but less sharp and without amplification. That might help explain many effects of implicit, or non-conscious, processing, as discussed further below. 3.5. Attention and working memory Selective attention has often been proposed to be one of the main functions of mechanisms that are referred to here as AA (e.g. Cauller, 1995; Jiang et al., 2013; M. Larkum, 2013; Larkum et al., 2004; Lee et al., 2014). Working memory (WM) is also often mentioned, and although that is clearly distinct from attention the two are closely related. Several researchers argue that attention and WM involve similar neural mechanisms because attention plays a leading role in the selection and maintenance of WM information (e.g. Awh & Jonides, 2001; Roth, Serences, & Courtney, 2006). Studies of visual WM and of interference with visualization showed long ago that initial and sustained attention are both major determinants of visual WM performance (e.g. Phillips & Christie, 1977; Phillips, 1983). Recent studies provide much further evidence for this, even where simple stimuli are to be remembered and where feature binding is required (e.g. Morey & Bieler, 2013). Prefrontal and parietal regions have a central role in generating these modulatory attentional influences. Functional MRI studies indicate that PFC is involved in both the attentional enhancement of relevant stimuli and the suppression of distracters (Gazzaley & Nobre, 2012). Furthermore, to test the causal role of

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

PFC-mediated modulation in attention and WM, repetitive transcranial magnetic stimulation has been used to temporarily impair function within specific sub-regions of PFC. This indicates a causal role for the PFC in the attentional modulation of neural activity in visual cortex during WM tasks (Zanto, Rubens, Thangavel, & Gazzaley, 2011). Overall, attention and working memory have much overlap in terms of both the sites of activity modulation in sensory and perceptual regions, and in sharing common sources for modulatory signals in prefrontal and parietal regions. These modulatory interactions can affect neuronal excitability both in the presence and in the absence of relevant external input (Gazzaley & Nobre, 2012). Computational, electrophysiological, and psychophysical studies showing that selective attention has modulatory effects are reviewed by Reynolds and Heeger (2009). Recent analyses of neuroimaging data suggest that attention modulates precision by top-down post-synaptic gain control (Brown & Friston, 2013). There are several hypotheses concerning the neuronal mechanisms by which such modulation is achieved (Rothenstein & Tsotsos, 2014). Most assume that it is at the microcircuit level, e.g. via a delicate balance between the background levels of excitatory and inhibitory synaptic currents (Salinas & Their, 2000). The intracellular capabilities reviewed in Section 2 now add another possibility, however, because they involve inputs from higher regions and more directly provide the modulatory effects that attention requires. The focus here is therefore on those few models of attention that have explored this possibility (e.g. Shipp, Adams, Moutoussis, & Zeki, 2009; Siegel et al., 2000; Spratling & Johnson, 2004). Siegel et al. (2000) were among the first to study computational capabilities that arise from networks built from neurons with functionally distinct inputs to their apical and basal dendrites. Their physiologically inspired computational study was of two-layer spiking networks built from neurons in which receptive field inputs formed synapses on the basal dendrites, and top-down attentional inputs formed synapses on the apical dendritic tree. Interactions between these two distinct sites of integration were modeled in accordance with intracellular evidence such as that provided by Larkum et al. (1999). Siegel et al. showed that the top-down input greatly enhanced response to the selected receptive field inputs. This enhancement consisted predominantly of an increase in burst firing, and was seen predominantly in the low-frequency components of response (<20 Hz). That finding was prescient because many neurophysiological studies now show that the effects of attention are indeed exerted predominantly at low beta frequencies, whereas feedforward influences are exerted predominantly at gamma and theta frequencies (Bastos et al., 2015; van Kerkoerle et al., 2014). Spratling and Johnson (2004, 2006) developed a computational model of the biased competition theory of attention in which the attentional bias was mediated via top-down input to the apical tufts. The top-down activation was used to modulate neuronal output by analogy to the findings of Larkum et al. (1999). In accordance with the intracellular evidence bottom-up input in this model could produce output even in the absence of top-down activity, whereas the amplifying feedback could not produce output in the absence of feedforward input. Their model faithfully replicated the selective and the facilitatory effects of spatial attention, as well as the effects of attention to particular features. In addition they showed that this form of modulation can also account for various other perceptual phenomena such as figureground segregation and contextual disambiguation. Their model has since been developed to account for many other psychophysical and cognitive phenomena (Spratling, 2008, 2010, 2012, 2014), and, though not in principle dependent on it, the top-down amplifying modulation in those later models could be also be implemented via AA (Spratling, personal communication).

7

Shipp et al. (2009) noted that layer 1 is the primary target of feedback from higher areas where it connects to the tufts of layer 2, 3, and 5 pyramidal cells. They then hypothesized that the basal dendrites of pyramidal cells specify their receptive field selectivity, and that attentional inputs to the apical tufts enhance the cell’s responsiveness, perhaps by a mechanism similar to BAC firing (e.g. Larkum et al., 2004). As an example of the functional consequences of such a mechanism they show in detail how it could produce attentive binding as in the cross-featural motion after-effect (Sohn, Papathomas, Blaser, & Vidnyanszky, 2004). 3.6. Setting the level of consciousness It has often been suggested that consciousness involves an interaction between feedforward and feedback signals, and this is at the center of the theory of consciousness proposed by Bachmann and Hudetz (2014). Put simply, their central hypothesis is that basal inputs specify the contents of consciousness and apical inputs determine the level of consciousness. They argue that basal input alone generates short-lived EPSPs and fragmented cortical neural activity but not a sustained ‘‘field of consciousness”. This corresponds to what Crick and Koch (2003) refer to as the ‘zombie’ mode of operation that is characterized by rapid stereotypical responses. If there is also sufficient modulatory input to the apical dendrites, however, then this amplifies neuronal output, thus producing the integrative computation experienced as being more conscious, flexible, and context-sensitive. Bachmann and Hudetz (2014) present several arguments in support of their hypothesis. First, anesthetics greatly reduce N1 and these effects are strongly correlated with the extent to which those anesthetics affect dendritic calcium spikes. Second, the effects of AA depend upon the presence of a feedforward signal to amplify. In the case of perception this ensures that the contents of consciousness are normally realistic rather than hallucinatory. Third, nonspecific thalamic input is thought to be necessary for consciousness, and that input is transmitted via layer 1, and thus via the tufts (M. Larkum, 2013). Fourth, consciousness operates on a 30–350 msec timescale, and this corresponds well to that on which BAC firing operates (Ledergerber & Larkum, 2012). Fifth, several varieties of visual masking and priming can be explained in detail by this hypothesis. Finally, they note that as BAC firing depends upon apical input from long-range inter-regional signals it is highly consistent with theories that propose such integration to be the neural basis of consciousness. Their hypothesis is further supported by a recent review concluding that many general anesthetics work by blocking the effects of predictive input received via apical dendrites (Meyer, 2015). Bachmann and Hudetz (2014) make a strong case for their hypothesis, but further clarifications are necessary. First, it can also be argued that consciousness depends upon the formation of transient coalitions (Crick & Koch, 2003), so the role of AA in creating large mutually-supportive coalitions needs to be clarified. Relating assembly formation to NMDA-dependent anesthesia may be one way of doing this (Flohr, Glade, & Motzko, 1998). Second, is it only the contents that are consciously experienced, but not the level, and, if so, why? Third, strong modulatory interactions can occur even in cases where the subject is not conscious of the contextual information producing the amplification (e.g. Troscianko et al., 1996). It therefore needs to be made clear how it is possible for effective modulation to occur without consciousness. Fourth, Bachmann and Hudetz assume that BAC firing is the mechanism by which apical input amplifies response, but the evidence reviewed in Section 2 suggests that other mechanisms may be more important in some classes of pyramidal cell. Fifth, Bachmann and Hudetz imply that amplification can be strong enough to amplify transmission of the content, but not strong enough to elevate it into

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

8

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

consciousness. What is the difference between the AA that is sufficient for consciousness and that which is not? Sixth, there is a long tradition in psychiatry, dating back to at least Janet, Freud and Jung, relating disorders such as hysteria and schizophrenia to impaired levels of consciousness (Silverstein, personal communication), so how is AA related to the evidence on which that tradition was based? Finally, some functions of AA, such as attention, working memory, and cognitive control require consciousness, whereas others, such as figure-ground segregation and priming, do not, so such differences require both conceptual and empirical clarification. 3.7. Cognitive control The cognitive control that PFC exerts over other regions is a form of contextual modulation with wide-ranging implications for normal and pathological function. As amplification is central to cognitive control (Egner & Hirsch, 2005), the question arises as to whether it involves AA. We know of no research concerned directly with this issue, but there are several reasons for supposing that AA has a major role. First, cognitive control amplifies relevant activities, so AA provides a possible mechanism for it (Phillips et al., 2015). Second, cognitive control signals are transmitted from PFC to the more posterior regions either directly or indirectly via nonspecific thalamus. In both cases the input from PFC to the receiving region is mediated by synapses in layer 1. Some of those synapses are on inhibitory interneurons located in layer 1 (Lee et al., 2014), but many of them are on the apical tufts of pyramidal neurons (Cauller & Connors, 1994; Rubio-Garrido et al., 2009). This implies that PFC operates via AA and its inhibitory regulation. Third, AA is dependent on NMDA receptors and there is evidence that they also have a central role in cognitive control. For example, the cognitive control of emotion is a crucial PFC function (Ochsner & Gross, 2005), and there is evidence that this control is highly sensitive to the detailed parameters of NMDAR-dependent transmission. For example, Bortolato et al. (2012) found that, as in humans, mice deficient in monoamine oxidase A are predisposed to abnormally aggressive outbursts in response to stress. This lack of control was associated with a change in the sub-unit composition of NMDA receptors such that NMDAR currents had reduced amplitude and an increased rate of decay. These findings were interpreted as indicating that reductions in excitability and increases in decay rate in NMDARs decrease the ability of PFC to integrate inputs arriving at different times from a wide variety of sources, thus impairing its ability to regulate emotional processing. All this suggests that much remains to be discovered concerning the role of AA in cognitive control. 3.8. Inter-regional coordination Siegel et al. (2000) provide a computational study of how higher cortical regions could usefully modulate the activity of lower regions via AA. Their model contained many biologically realistic details including two distinct classes of inhibitory interneuron, and predicts that if this form of AA occurs in the cortex it will increase bursting more than it increases firing rates. It also predicts that feedback will be transmitted mainly in the lower frequency bands (<20 Hz), as indicated by recent findings (e.g. Bastos et al., 2015; Boudewijns et al., 2013; van Kerkoerle et al., 2014). It is well-established that, in addition to feedforward and feedback pathways, there are lateral pathways linking local circuits that are at the same hierarchical level. For example, the ventral pathway recognizes objects where-ever they are; the dorsal pathway locates them what-ever they are. Thus, signal for one pathway is noise for the other, and AA provides a mechanism by which

separate pathways can modulate each others’ activities while preserving their distinctive receptive field selectivities. 3.9. Learning If AA is closely related to attention and/or consciousness then that implies that it will also be closely related to learning, which depends to a large extent on attention and consciousness. AA is likely to occur when pyramidal cells produce action potentials that are part of a temporary coalition as proposed by Crick and Koch (2003), because such coalitions are formed when the information transmitted by each cell is coherently related to the broader context of current activity elsewhere. As that information is received via the apical tuft AA is likely to play a major role in learning. This hypothesis fits well with what is known about the synaptic plasticity that underlies learning. Long-term potentiation has been shown to be sensitive to both calcium conductances and bursting (e.g. Pike, Meredith, Olding, & Paulsen, 1999), and BAC firing converts a single action potential into a brief burst via calcium spikes. Körding and König (2000) investigated the properties and capabilities of a neurobiologically inspired computational model of learning using neurons with two sites of synaptic integration. Their model showed how, among other learning capabilities, simulated networks built from elements with distinct basal and apical sites of integration can learn representations that are translation invariant (Einhauser, Kayser, Körding, & König, 2002; Körding & König, 2000). A series of other computational studies that are conceptually close to those of Körding & König, but more abstract, also examined the learning capabilities of neural networks built from local processors with distinct sites for integrating contextual and receptive field information (Kay & Phillips, 2011; Kay et al., 1998; Phillips et al., 1995). Those studies show that such networks have extensive learning capabilities, including the ability to discover latent statistical structure without requiring any supervision. None of this implies that AA is necessary for learning, however, and it is not easy to see how it could be because AA seems to arise only at the later stages of phylogenetic and ontogenetic development. Thus, much remains to be discovered concerning the conditions under which AA does and does not play a role in learning. 3.10. Abnormalities of apical amplification Much light may be cast on the role of AA in cognition by studies of cognitive impairment due to various forms of AA-malfunction. Many different patterns of cognitive impairment are possible depending on exactly how AA or disamplification are impaired, when, to what extent, in which cognitive sub-systems, and in combination with which other abnormalities. Such abnormalities could involve under-amplification, over-amplification, or various forms of unreliable amplification. Great heterogeneity can therefore be expected in those abnormalities. A few are briefly listed here. Though none prove that AA occurs or has the functions hypothesized, they do provide supportive evidence and suggest many issues worthy of much closer examination for clinical as well as for basic scientific reasons. Though the anesthetized state is not an illness it is an abnormal state that may provide valuable insight into neural mechanisms because it is due to known and controllable causes. Calcium spikes in the apical dendrites of layer 5 neurons are greatly reduced by anesthetics (Potez & Larkum, 2008) and rapidly return to normal after release from the anesthetic (Murayama & Larkum, 2009). Furthermore, there is good evidence that a wide variety of anesthetics operate via the reduction of AA (Meyer, 2015). Such findings have therefore been used as a major source of support

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

for the hypothesis that setting the level of consciousness is a major function of BAC firing (Bachmann & Hudetz, 2014). Neither BAC firing nor AA in general can be sufficient for consciousness, however, because much of the intracellular evidence for them comes from in vitro slices or from anesthetized animals. Furthermore, we need to explain why attention, working memory, and cognitive control are greatly impaired by the loss of consciousness in anesthesia, whereas other functions of AA are not. Ketamine anesthesia does not affect inhibitory interneurons but works by inhibiting AA, and comes close to dissociating wakefulness from consciousness, because, although unresponsive to their environment, anesthetized patients may keep their eyes open and later recount sensory distortions and hallucinations (Meyer, 2015). Maybe ketamine blocks some functions of AA but not others depending on the dose. In-depth reviews show that schizophrenia involves impairments in all of the cognitive functions listed above, together with malfunctions of the NMDARs on which AA depends (Adams, Stephan, Brown, Frith, & Friston, 2013; Phillips & Silverstein, 2003, 2013; Phillips et al., 2015). Compelling evidence for an association between schizophrenic symptoms and NMDARs comes from autoimmune anti-NMDAR encephalitis, which progressively reduces activity at NMDA receptors by capping and internalizing them (Hughes et al., 2010). These patients present with symptoms that are so like schizophrenia that, prior to its further progression, schizophrenia is often the initial diagnosis. As basal dendrites also have NMDARs the hypothesis of a relation between their malfunction and AA suggests that NMDARs somehow play a more crucial role in apical dendrites than they do in basal dendrites. There is good support for that implication. NMDAR antagonists abolish calcium transients in apical dendrites (Palmer et al., 2014), whereas in basal dendrites they linearize response but do not abolish it (Branco & Häusser, 2011; Schiller, Major, Koester, & Schiller, 2000). Linearization of integration in the basal dendrites will degrade their normal operation by removing non-linear aspects of their integration (Jadi et al., 2014) but basal inputs will nevertheless still be able to drive output. Thus this shows how the extensive evidence relating NMDAR hypofunction to schizophrenia (Phillips & Silverstein, 2003, 2013; Moghaddam & Javitt, 2012) can be reconciled with both the specificity of schizophrenic symptoms and the ubiquity of NMDARs. Abnormalities of AA may also be involved in autism, which also involves aberrant settings of precision (Lawson, Rees, & Friston, 2014), and thus of contextual modulation. Intracellular studies of pyramidal cell physiology in autistic and fragile-X disorders suggest, that, in contrast to schizophrenia, they involve overamplification or unreliable amplification rather than underamplification (Markram & Markram, 2010; Meredith, Holmgren, Weidum, Burnashev, & Mansvelder, 2007; Zhang et al., 2014). Autistic disorders involve abnormalities that become serious much earlier in the course of development than do schizophrenic abnormalities, so cognitive development will be much more affected. Ways in which it is affected might therefore reflect aspects of cognitive development that depend upon the contextual guidance provided by AA. Relevant aspects are suggested by the domain-general interactive specialization framework for development (Johnson, 2011) that draws upon evidence for BAC firing as modeled by Siegel et al. (2000). Detailed examination of the relation between those aspects of development and autistic spectrum disorders is a major task for the future. Some developmental aspects of apical function have already been studied using slices of rat neocortex (Atkinson & Williams, 2009). They show that apical inputs that powerfully excite layer 5 cells at early post-natal stages have a progressively weaker impact as animals mature. This is accompanied by an age-dependent switch in the integrative operations of pyramidal neurons, from one analogous to a

9

single-compartment integrate-and-fire point-processor early in postnatal development to a processor with distinct somatic and apical sites of integration later in development. Another form of AA-malfunction is due to the effects of exposure to alcohol early in development. An anatomical study of the spine density of layer 2 and 3 pyramidal neurons in the somatosensory cortex of rats exposed to alcohol during the first postnatal week found that it decreased adult spine density in the apical dendrites, but had no such effect on the basal dendrites (De Giorgio & Granato, 2015). Early exposure to alcohol also has lasting effects on the excitability of layer 5 cells due to down-regulation of calcium electrogenesis (Granato, Palmer, De Giorgio, Tavian, & Larkum, 2012). Thus, if alcohol impairs AA, and if its functions are as listed above then those functions should be impaired in fetal alcohol spectrum disorders. Current evidence clearly indicates that they are (Guerri, Bazinet, & Riley, 2009; Phillips, in press), so this issue merits much further study. 3.11. Implications for theories of cognition based on integrate-and-fire point neurons Most theories of the neuronal bases of cognition assume that neurons can be adequately represented as ‘point processors’ that integrate all of their excitatory and inhibitory synaptic currents at the soma, and use that single integrated value to determine output firing rate. There are strong grounds for doubting the validity of that assumption (Jadi et al., 2014). Though long questioned (e.g. Koch & Segev, 2000), it remains widely accepted. The evidence and arguments implicating AA in a wide range of basic cognitive functions show that we should now advance beyond it. 4. The relevance of apical amplification to potentially unifying theories of cortical computation As mechanisms for AA are widely distributed throughout the cortex, and as the list of specific cognitive functions proposed for it is broad, it should be possible to describe its function in a more abstract, domain-general, or ‘unifying’ way (Wibral et al., this issue). Theories offering such a description are therefore briefly discussed here in relation to the evidence for AA. 4.1. Coherent infomax and related theories The theory of coherent infomax was developed in the early days of connectionism and machine learning, and used a formally specified objective to interpret evidence on the structure and function of neocortex. It provides an information theoretic account of both learning and processing by formally specifying their objective as maximizing the amount of coherent information transmitted (Kay & Phillips, 1997; Phillips et al., 1995), which implies increasing the organism’s organized complexity (Phillips, 2012). The learning rule derived analytically from this objective discovers latent statistical structure in large data sets by adjusting synaptic strengths. In broad outline this learning rule has biological plausibility (Phillips & Singer, 1997), and approximations exist by which it can be adapted to neural networks with driving and modulatory input vectors of high dimensionality (Kay & Phillips, 2011). The theory shows how the knowledge thus acquired can amplify coherent signals and suppress incoherent signals. It is potentially ‘unifying’ in the sense that it specifies a highly abstract goal toward which neural systems evolve across a wide-range of timescales. Though many sources of evidence were used to guide the development of the theory of coherent infomax it may nevertheless seem to be too abstract and grandiose to be plausible. That is

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

10

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

why, from the perspective of that theory, the evidence for AA seems to be a major empirical breakthrough. It provides an intracellular basis for exacting computations that the theory requires, and it does so by mechanisms that seem to be common to widespread classes of pyramidal cell. That clearly encourages the view that AA and disamplification are widespread throughout the neocortex, so that may qualify them as being neurocomputational primitives together with the long-established primitives of excitation, inhibition, and disinhibition. It is in that sense that the discovery of AA is claimed to be ‘unifying’. Like those other primitives, AA and disamplification seem to be involved in a wide range of behaviors. Though these five primitives do not provide a complete explanation of any of those behaviors, they do provide elementary computational capabilities of relevance to them all. The infomax theory developed by Linsker (1988) derived many properties of visual cortex from the assumption that the objective of each of the many local processors of which it is composed is to maximize the mutual information between their inputs and outputs under the strong constraint of data reduction. This requires the discovery of latent statistical structure and leads to efficient coding or redundancy reduction. Coherent infomax is a broader theory that includes efficient coding as a sub-goal. It derives general properties of neocortex from the assumption that a fundamental objective of brain function is to maximize the transmission of information that is coherently related to the context within which it occurs, where contextual information is provided by modulatory inputs to the local processors. Coherent infomax is therefore fundamentally different from infomax because its objective relates three distinct sets of variables, not just two. Surprisingly, extending the information theoretic analysis rigorously from two to three sets of variables requires recent advances in information theory that are not yet widely known (Wibral et al., this issue). Reviews and developments of coherent infomax are presented elsewhere (e.g. Kay, 2000; Kay & Phillips, 2011; Phillips, 2012; Phillips et al., 2015). The most relevant aspect here is that the learning rules must be formulated in a way that is feasible within networks composed of many local processors, each with many inputs. This requires their driving and modulatory inputs to come from different sources, and to be integrated separately (Kay & Phillips, 2011). It has been argued that these capabilities are provided by NMDA receptors because they are clearly implicated in modulatory functions (Phillips & Silverstein, 2003). That argument faces one great difficulty, however. NMDA and AMPA receptors are co-localized on the spines of both basal and apical dendrites so they cannot receive their inputs from different sources, and those inputs cannot first be integrated separately. The discovery of AA solves that problem because apical inputs come from sources that are clearly distinct from those of the basal inputs, and they are integrated separately before being used to modulate response to the basal inputs; i.e. AA and its disamplification provide exactly the capabilities that the theory of coherent infomax requires. Furthermore, it does so via NMDA-dependent mechanisms! The learning rule of coherent infomax has essentially the same objective as that in the learning theory of ‘relevant infomax’ (Körding & König, 2000), and the short-term dynamics are essentially the same as those studied by Siegel et al. (2000) which is based explicitly on AA. A rigorous computational framework developed by Spratling (2002, 2008) and by Spratling and Johnson (2004, 2006) offers a similar theory that has so far been related in most detail to vision and attention, but which has the potential to cover many of the cognitive functions discussed above. All of these theories emphasize the potential importance for cognition of the physiological evidence for two intracellular sites of integration, with one being modulatory.

4.2. Counter-streams The ‘counter-stream’ theory has ascending streams of processing originating from external input and descending streams originating from stored models (Ullman, 1995). These two streams are assumed to correspond to feedforward and feedback connections between cortical areas. The theory has four key ingredients. First, many ascending and descending streams operate concurrently across several hierarchical levels. Second, separate nodes (i.e. coactive populations of neurons) integrate ascending and descending signals, but are paired such that there is a one-to-one relationship between nodes in ascending and descending streams. Third, ascending nodes are driven by ascending activity, and descending nodes by descending activity. Activities with internal and external origins are thus strictly separated. Fourth, when any unit receives activation it sends a signal to its paired partner which ‘primes’, or amplifies, response of the partner to its own driving signals. Thus sequences of mappings across different levels of the hierarchy become more highly activated when ascending and descending activations match. Ullman argued that these ‘priming’, or amplifying, interactions provide a basis for the contextsensitivity observed in many psychological studies. The approach was assumed to have generic utility, and computational simulations showed that in visual pattern recognition tasks it performed as well or better than the alternatives then available. The theory was shown to have neurobiological plausibility, and has recently received further support by anatomical studies showing that there are two topographically organized feedforward and feedback pathways, one in layers 2 and 3, and one in layers 5 and 6 (Markov et al., 2014). The counter-stream in layers 2 and 3 allows fine detailed information to travel both up and down the cortical hierarchy as the theory predicts. Modulatory interactions are crucial to the theory. It predicts that corresponding nodes in the ascending and descending streams will have strictly segregated long-range inputs, and reciprocal local connections that are amplifying but not driving. It is now known that in macaque visual cortex descending inputs to layers 2/3A are strictly segregated from ascending inputs to layers 4 and 3B (Markov et al., 2014). Ullman (1995) hypothesized that interactions between ascending and descending streams may be mediated by apical tufts. Markov et al. (2014) confirmed that in adult macaque visual cortex neurons in layers 2 and 3, whether feedforward or feedback, had a tufted apical dendrite. They summarize their findings as showing that feedforward neurons of layers 3B and 5 interact with feedback signals via their apical dendrites in layer 1. This echoes findings in mouse somatosensory cortex (Petreanu et al., 2009). The opportunity for such counter-streams to interact via AA therefore clearly exists. We now need to know the conditions under which active dendritic currents in the apical trunk make use of that opportunity. As Markov et al. (2014) conclude, the counter-stream hypothesis has several strengths, but it also has some weaknesses. First, the anatomical data show that there is far from a one-to-one correspondence between cells in the two streams; feedback cells span greater cortical distances and are more numerous. Second, though there is evidence that some pyramidal cells can be driven by feedback, most are driven by feedforward input. Third, much modulatory input affects feedforward activity directly rather than being conveyed indirectly via a neuron driven by descending input. Fourth, it would be inefficient to implement lateral modulatory interactions, such as those that occur plentifully both within and between regions, via paired nodes, and there is no evidence that it is. These difficulties do not negate the merits of the counterstream theory, however, which, in agreement with the theories discussed in Section 4.1, implies that cognition depends upon local cortical circuits that receive modulatory inputs that are strictly segregated from their driving inputs.

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

4.3. The Bayesian brain, predictive coding and free energy minimization This group of theories is too large to be summarized in a few paragraphs. Although they have not yet achieved a widely accepted consensus, they have been extensively developed and are highly influential, as other papers in this special issue demonstrate. Amongst the most important things offered by these theories is the prospect of a deep unification that shows fundamental commonalities between perceptual and motor processing (Friston, 2010) while explaining ways in which they differ (Shipp, Adams, & Friston, 2013). It may also show deep commonalities between basic cognitive processes and fundamental aspects of higher cognitive functions such as self-identity, body ownership, and the sense of agency (Seth, 2013; Seth et al., 2011). These theories have not yet been related in any detail to the evidence for AA, and the following comments raise important issues that are as yet unresolved. A few comments will first be made on relations between AA and Bayesian inference. First, Bayesian inference does not require AA. It can be implemented in neural networks that use interactions that are multiplicative but not modulatory (e.g. George & Hawkins, 2009). The distinction between multiplicative and modulatory interactions is crucial but not well-known. Multiplicative interactions are symmetrical, modulatory interactions are not. From the perspective of coherent infomax the role of contextual modulation in probabilistic inference is fundamentally dependent on this asymmetry (Kay & Phillips, 2011; Phillips, 2012). It is clear in the evidence for AA because that shows that the functions of the two sites of integration are not equivalent; one directly generates action potential output, the other modulates response without directly generating output. The distinction between modulatory and multiplicative interactions implied by this argument is further supported by much evidence showing that many forms of modulatory interaction are impaired in schizophrenia, but coordinate transformations, which require multiplicative but not modulatory interactions, are unimpaired (Phillips & Silverstein, 2013; Phillips et al., 2015). Second, the evidence for AA does not imply that the brain is Bayesian. For example, it could implement attentional and many other cognitive processes without any use of Bayesian inference (Spratling, 2002; Spratling & Johnson, 2004, 2006). Third, some of the inputs to the apical integration site convey signals from other local circuits that are statistically correlated with activity in the target circuit. Thus, those inputs provide additional information that could be used in deciding how to respond to any current driving inputs. Fourth, contextual modulation must operate, not via priors as is widely assumed, but via likelihoods. This is because priors must be specified prior to knowing the data, whereas modulatory effects depend upon knowing the data (Kay & Phillips, 2011). Fifth, the heuristic of thinking of cells and circuits as explicitly doing mathematical calculations may be misleading if taken too literally. A neurocentric Jaynesian perspective may therefore be of more use to neuroscience than the conventional Bayesian perspective (Fiorillo, 2010, 2012; Phillips, 2012). From that perspective there is no need for neural systems to explicitly perform any of the mathematics of probability theory (Fiorillo, 2012). Sixth, it can be argued that Bayesian inference and predictive coding are primarily concerned with coding of the information to be transmitted (Fiorillo, 2012). In contrast to that, AA seems to be more concerned with selecting the information to be transmitted. Finally, some inputs may amplify response to feedforward activity, not because they were or were not predicted, but because they have special relevance as specified by attentional or evaluative processes. Several comments can be made concerning relations between AA and theories of predictive coding and free energy reduction.

11

First, there is disagreement concerning relations between these theories and AA as modeled by Körding and König (2000) and by Siegel et al. (2000). Those models make no use of ‘error units’, and Siegel et al. (2000) note that their model contrasts with predictive coding models because theirs depends upon top-down interactions that are excitatory, whereas in predictive coding models they are inhibitory. Kveraga et al. (2007) take a different view, however, and claim that the underlying neural architecture of Siegel et al. (2000) could be adapted to fit with predictive coding theory, the ‘counter-stream’ theory (Ullman, 1995), the theory of free energy reduction (Friston & Stephan, 2007), and adaptive resonance theory (Grossberg, 1980). Kveraga et al.’s claim therefore implies that all those theories could make use of AA. The validity of that claim remains to be established. Second, one of the things that the preceding issue depends on is the concept of ‘prediction error’. If whatever is fed forward is prediction error, as suggested by Shipp et al. (2013) using complex cells in V1 as an example, then AA could provide a mechanism for amplifying prediction error signals, because layer 3 pyramidal cells, which are a major source of inter-regional feedforward signals, have extensive tufts in layer 1 (e.g. Gur & Snodderly, 2008; Petreanu et al., 2009). Note, however, that such evidence does not show that it is useful to describe whatever is fed forward as ‘prediction error’. What it shows is that input to the tuft, wherever it comes from, amplifies feedforward signals, whatever they code for. Third, it can be assumed that all neural signals are usually informative, implying that they will usually reduce uncertainty whatever their source. Furthermore, given that more abstract ‘features’ or ‘concepts’ can be instantiated in multiple ways at lower levels, feedforward signals will usually be informative because they specify the particular way in which higher-level sensory features or abstract concepts are presently instantiated at lower hierarchical levels. Fourth, there may be a central role for AA in free energy theory because it could provide an intracellular mechanism for the control of gain, precision, or confidence that is central to the theory (Brown, Adams, Parees, Edwards, & Friston, 2013; Clark, 2013a, 2013b; Friston, 2010; Friston, Bastos, Pinotsis, & Litvak, 2015). From its early days free energy theory has been primarily concerned with analyzing the macroscopic dynamics of whole neural systems adapting to and operating within an environment. Therefore, as Friston et al. (2015) say, a long road ahead faces attempts to test and develop it in terms of intracellular processes. Consideration of the evidence for AA may provide some early steps along that road, and supports the free energy theory in several ways. It shows that contextual modulation is widespread throughout the cortex and plays a central role in regulating its dynamics. It supports the view that feedback signals are functionally distinct from feedforward signals, have an important role in modulation, and can directly amplify the transmission of feedforward signals (e.g. as hypothesized by Shipp et al., 2013). It also supports the view that NMDA receptors have a key role in modulation because AA has been shown to be NMDAR-dependent. Finally, neuroimaging evidence indicates that gamma frequencies predominate in feedforward signals, whereas lower frequencies predominate in feedback signals as predicted by free energy theory (Bastos et al., 2012). AA, as modeled by Siegel et al. (2000), resonates with that because it also produces similar differences in the spectral composition of feedforward and feedback signals. Fifth, there are also ways in which the evidence for AA raises difficulties for the free energy and predictive coding theories. One obvious difficulty is that within free energy and predictive coding theories pyramidal cells are usually treated as point processors with a single site of integration. Shipp et al. (2013) do suggest that input to apical tufts may be used for modulatory control of

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

12

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

precision in the case of layer 3B cells, but the evidence for AA suggests that the sub-set of pyramidal cells with two distinct sites of integration is far larger than that, also including cells of layers 2 and 5. Furthermore, the tufts of neurons in layers 2, 3 and 5 are intermingled in layer 1, suggesting that they may share a common amplifying input. Another difficulty is that the evidence for AA shows clearly that input to the apical tufts comes from a diverse range of sources and is not limited just to feedback from the succeeding hierarchical level (Gur & Snodderly, 2008). As feedback via layer 1 is combined with information from such a diverse variety of other sources it is hard to see how it can determine both what the feedforward signals code for and their strength or precision. Shipp et al. (2013) suggest that feedback from layer 6 of down-stream regions may play the role of descending expectations in determining what feedforward signals code for, but that suggestion remains to be validated. Prima facie it seems unlikely that the channel capacity of feedback from layer 6 cells is sufficient to predict more than a small fraction of the ascending information. Furthermore, if layer 6 cells have essentially the same active apical dendritic physiology as layer 5 cells, then they receive their apical inputs from the feedforward inputs to layer 4 rather than from the diverse inputs to layer 1, with functional consequences that remain to be elucidated. Yet another difficulty is that the signature of a successful match between feedforward data and expectation in predictive coding and free energy theories is silence. In contrast to that, the signature of a successful match between apical and basal inputs is the generation and synchronization of bursts containing low inter-spike intervals in a sparse subset of cells. Another difficulty is that feedback could reduce total activity in a region by focussing it on a small sub-set of highly selected columns via amplification and disamplification but without changing coding. Such effects would therefore need to be distinguished from any effects of feedback on coding, posing intriguing questions for future empirical and theoretical explorations. Furthermore, inputs to the tuft could have large effects on dendritic currents while having little effect on output. Some of the findings used to support free energy and predictive coding theories (e.g. by Muckli & Petro, 2013) may therefore reflect the focusing of action-potential activity on a few columns or it may reflect dendritic activity in the tuft rather than effects upon action potential output (Cauller, 1995), thus raising further ambiguities that need to be resolved.

5. Questions for further research The evidence and arguments reviewed suggests that AA is a fundamental neurocomputational capability with relevance to several different cognitive domains. If so, there is a rich story to be told concerning its evolution, development, regulation, functions and malfunctions. As attempts to read that story are still young, however, many unresolved issues arise. In particular much remains to be discovered concerning the conditions under which AA occurs during normal awake conditions. Figs. 1 and 2 given in the introduction could be misleading because they do not change over time. It is likely that AA is much less fixed, however, because the active dendritic currents by which it is implemented depend upon many things. It is not yet firmly established that AA occurs in layer 2 and 3 cells in awake behaving animals. Furthermore, the extent to which AA depends on calcium spiking is not yet clear because apical calcium channels may influence action-potential output without spiking. Observations made on slices or anesthetized animals may make AA look either more or less important than it actually is. It seems likely that those observations underrepresent its importance given the circumstantial evidence and theoretical arguments presented above. There is already some direct evidence that AA may be more effective in the awake than

in the anesthetized state (Murayama & Larkum, 2009), as implied by the arguments and evidence linking it to consciousness. The dependence of AA on current mental state variables, such as those controlling the level of arousal both tonically and phasically, is therefore a crucial issue to be studied. Many other issues arise. Some are briefly noted here, along with references to a few key papers that provide relevant evidence or discussion. ERP data has been interpreted as providing evidence on the role of apical dendritic currents in human cognition (Cauller, 1995), but how strong is that evidence, and how can it be strengthened? What is the relevance of AA for the analysis of other macroscopic measures? How do the NMDA-dependent intracellular mechanisms for contextual modulation evolve (Cook, Carvalho, & Damasio, 2014; Ryan et al., 2013; Shepherd, 2011)? How do they develop (Atkinson & Williams, 2009)? What advantages does AA have over amplification by disinhibition (Spratling, 2002, 2008)? Is AA both signalled and generated by brief bursts of 2–4 action potentials within 10–20 msec (Boudewijns et al., 2013; Siegel et al., 2000)? In contrast to much other evidence, Anderson, Mitchell, and Reynolds (2013) report a paradigm in which attention does not increase bursting, but may even reduce it! Though bursting is not a sine qua non of contextual modulation as defined within the theory of coherent infomax, the findings of Anderson et al need to be reconciled with the assumption that AA provides a mechanism for attentional enhancement. Does some form of AA also occur in layer 6 cells? They have an active dendritic physiology that seems similar to that of other neocortical pyramidal cells, but with their apical integration site receiving inputs via layer 4 rather than via layer 1 (Ledergerber & Larkum, 2010; Qi & Feldmeyer, 2015). So, what are the functional implications of those differences? Answers to such questions will help us advance beyond long-held assumptions concerning the computational capabilities of individual neurons. Since McCulloch and Pitts (1943) proved that basic logical operations can in principle be implemented by networks composed of simple binary units with only excitatory and inhibitory inputs, nearly all theories in cognition and neuroscience have been based on such integrate-and-fire point processors (Jadi et al., 2014). If networks of sufficient complexity can in principle compute anything computable using only such simple elements then why do some neural systems use elements that are far more complex? They may do so because what could be done ‘in principle’ with simple elements given unlimited time and storage capacity is irrelevant to survival, which requires fast, effective, and efficient actions that are flexibly adapted to current circumstances using context-sensitive modulatory mechanisms. Perhaps direct monosynaptic amplifying interactions between pyramidal cells evolved because, in comparison to amplification by disinhibition, they are faster, more efficient, more locally specific, and more adaptable. If it is indeed the case that apical amplification and disamplification are fundamental primitives of neocortical computation then many decades of future research will be needed to explore their distribution, variations, capabilities, development, and evolution. Acknowledgments Several psychologists, neurophysiologists, neuroanatomists, clinicians, and theoretical neuroscientists in Stirling, Frankfurt, and beyond have contributed to the conceptual framework further developed here. The review of the evidence for apical amplification was greatly enhanced by the many suggestions made by Matthew Larkum. Comments leading to substantial improvements of the paper were also provided by: Michael Spratling, Alberto Granato, Heiko Neumann, Talis Bachmann, Christopher Fiorillo, André Bastos, Michael Wibral, Dimitris Pinotsis, Peter König, Steve

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

Silverstein, and Rosana Tristao. I also thank two anonymous reviewers for insightful comments that were used to improve the paper. References Adams, R. A., Stephan, K. E., Brown, H. R., Frith, C. D., & Friston, K. J. (2013). The computational anatomy of psychosis. Frontiers in Psychiatry, 4. http://dx.doi.org/ 10.3389/fpsyt.2013.00047. Allison, T., Puce, A., & McCarthy, G. (2002). Category-sensitive excitatory and inhibitory processes in human extrastriate cortex. Journal of Neurophysiology, 88, 2864–2868. Anderson, E. B., Mitchell, J. F., & Reynolds, J. H. (2013). Attention-dependent reductions in burstiness and action-potential height in macaque area V4. Nature Neuroscience, 16(8), 1125–1131. Atkinson, S. E., & Williams, S. R. (2009). Postnatal development of dendritic synaptic integration in rat neocortical pyramidal neurons. Journal of Neurophysiology, 102, 735–751. Awh, E., & Jonides, J. (2001). Overlapping mechanisms of attention and spatial working memory. Trends in Cognitive Sciences, 5(3), 119–126. Bachmann, T., & Hudetz, A. G. (2014). It is time to combine the two main traditions in the research on the neural correlates of consciousness: C = L  D. Frontiers in Psychology, 5. http://dx.doi.org/10.3389/fpsyg.2014.00940. Bastos, A. M., Usrey, W. M., Adams, R. A., Mangun, G. R., Fries, P., & Friston, K. J. (2012). Canonical microcircuits for predictive coding. Neuron, 76(4), 695–711. Bastos, A. M., Vezoli, J., Bosman, C. A., Schoffelen, J.-M., Oostenveld, R., Dowdall, J. R., ... Fries, P. (2015). Visual areas exert feedforward and feedback influences through distinct frequency channels. Neuron, 85, 1–12. http://dx.doi.org/ 10.1016/j.neuron.2014.12.018. Bathellier, B., Margrie, T. W., & Larkum, M. E. (2009). Properties of piriform cortex pyramidal cell dendrites: Implications for olfactory circuit design. Journal of Neuroscience, 29(40), 12641–12652. Bayerl, P., & Neumann, H. (2004). Disambiguating visual motion through contextual feedback modulation. Neural Computation, 16, 2041–2066. Bortolato, M., Godar, S. C., Melis, M., Soggiu, A., Roncada, P., Casu, A., ... Shih, J. C. (2012). NMDARs mediate the role of monoamine oxidase A in pathological aggression. Journal of Neuroscience, 32, 8574–8582. Boudewijns, Z. S. R. M., Groen, R. M., Lodder, B., Minni, T. B., McMaster, M. T. B., Kalogreades, L., ... de Kock, C. P. J. (2013). Layer-specific high-frequency action potential spiking in the prefrontal cortex of awake rats. Frontiers in Cellular Neuroscience, 7, 99. http://dx.doi.org/10.3389/fncel.2013.00099. Branco, T., & Häusser, M. (2011). Synaptic integration gradients in single cortical pyramidal cell dendrites. Neuron, 69, 885–892. Brosch, T., & Neumann, H. (2014). Computing with a canonical neural circuits model with pool normalization and modulating feedback. Neural Computation, 26, 1–55. Brown, H., Adams, R. A., Parees, I., Edwards, M., & Friston, K. (2013). Active inference, sensory attenuation and illusions. Cognitive Processing, 14, 411–427. http://dx. doi.org/10.1007/s10339-013-0571-3. Brown, H. R., & Friston, K. J. (2013). The functional anatomy of attention: A DCM study. Frontiers in Human Neuroscience, 7. http://dx.doi.org/10.3389/ fnhum.2013.00784. Budd, J. M. (1998). Extrastriate feedback to primary visual cortex in primates: A quantitative analysis of connectivity. Proceedings of the Royal Society of London B: Biological Sciences, 265, 1037–1044. http://dx.doi.org/10.1098/ rspb.1998.0396. Cauller, L. J. (1995). Layer I of primary sensory neocortex: Where topdown converges upon bottom-up. Behavioural Brain Research, 71, 163–170. Cauller, L. J., & Connors, B. W. (1994). Synaptic physiology of horizontal afferents to layer I in slices of rat SI neocortex. Journal of Neuroscience, 14(2), 751–762. Clark, A. (2013a). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36, 131–204. Clark, A. (2013b). The many faces of precision. Frontiers in Psychology, 4, 270. http:// dx.doi.org/10.3389/fpsyg.2013.00270. Clowes, M. B. (1971). On seeing things. Artificial Intelligence, 2, 79–116. Cook, N. D., Carvalho, G. B., & Damasio, A. (2014). From membrane excitability to metazoan psychology. Trends in Neurosciences, 37, 698–705. Crick, F., & Koch, C. (2003). A framework for consciousness. Nature Neuroscience, 6, 119–126. Davenport, J. L., & Potter, M. C. (2004). Scene consistency in object and background perception. Psychological Science, 15, 559–564. De Giorgio, A., & Granato, A. (2015). Reduced density of dendritic spines in pyramidal neurons of rats exposed to alcohol during early postnatal life. International Journal of Developmental Neuroscience, 41, 74–79. De Kock, C. P. J., Bruno, R. M., Spors, H., & Sakmann, B. (2007). Layer-and cell-typespecific suprathreshold stimulus representation in rat primary somatosensory cortex. Journal of Physiology, 581, 139–154. DeLong, K. A., Urbach, T. P., & Kutas, M. (2005). Probabilistic word pre-activation during language comprehension inferred from electrical brain activity. Nature Neuroscience, 8, 1117–1121. Duffy, S. A. (1986). Role of expectations in sentence integration. Journal of Experimental Psychology: Learning, Memory, and Cognition, 12, 208–219. Egner, T., & Hirsch, J. (2005). Cognitive control mechanisms resolve conflict through cortical amplification of task-relevant information. Nature Neuroscience, 8(12), 1784–1790.

13

Einhauser, W., Kayser, C., Körding, K. P., & König, P. (2002). Learning the invariance properties of complex cells from their responses to natural stimuli. European Journal of Neuroscience, 15(3), 475–486. Fedota, J. R., McDonald, C. G., Roberts, D. M., & Parasuraman, R. (2012). Contextual task difficulty modulates stimulus discrimination: Electrophysiological evidence for interaction between sensory and executive processes. Psychophysiology, 49, 1384–1393. Fiorillo, C. D. (2010). A neurocentric approach to Bayesian inference. Nature Reviews Neuroscience, 11(8), 605. Fiorillo, C. D. (2012). On the need for a unified and Jaynesian definition of probability and information within neuroscience. Information, 3, 175–203. Flohr, H., Glade, U., & Motzko, D. (1998). The role of the NMDA synapse in general anesthesia. Toxicology Letters, 100, 23–29. Friston, K. J. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11, 127–138. Friston, K. J., Bastos, A. M., Pinotsis, D., & Litvak, V. (2015). LFP and oscillations— What do they tell us? Current Opinion in Neurobiology, 31, 1–6. Friston, K. J., & Stephan, K. E. (2007). Free energy and the brain. Synthese, 159, 417–458. http://dx.doi.org/10.1007/s11229-007-9237-y. Gazzaley, A., & Nobre, A. C. (2012). Top-down modulation: Bridging selective attention and working memory. Trends in Cognitive Sciences, 16(2), 129–135. Gentet, L. J. (2012). Functional diversity of supragranular GABAergic neurons in the barrel cortex. Frontiers in Neural Circuits, 6, 1–13. http://dx.doi.org/10.3389/ fncir.2012.00052. George, D., & Hawkins, J. (2009). Towards a mathematical theory of cortical microcircuits. PLoS Computational Biology, 5(10), e1000532. Ghuman, A. S., Bar, M., Dobbins, I. G., & Schnyer, D. M. (2008). The effects of priming on frontal–temporal communication. Proceedings of the National Academy of Sciences of the United States of America, 105, 8405–8409. Gilbert, C. D., & Sigman, M. (2007). Brain states: Top-down influences in sensory processing. Neuron, 54, 677–696. Granato, A., Palmer, L. M., De Giorgio, A., Tavian, D., & Larkum, M. E. (2012). Early exposure to alcohol leads to permanent impairment of dendritic excitability in neocortical pyramidal neurons. Journal of Neuroscience, 32, 1377–1382. Grossberg, S. (1980). How does a brain build a cognitive code? Psychological Review, 87, 1–51. Guerri, C., Bazinet, A., & Riley, E. P. (2009). Foetal alcohol spectrum disorders and alterations in brain and behaviour. Alcohol and Alcoholism. http://dx.doi.org/ 10.1093/alcalc/agn105. Gur, M., & Snodderly, D. M. (2008). Physiological differences between neurons in layer 2 and layer 3 of primary visual cortex (V1) of alert macaque monkeys. Journal of Physiology, 586, 2293–2306. Harris, K. D., & Shepherd, G. M. (2015). The neocortical circuit: Themes and variations. Nature Neuroscience, 18, 170–181. Häusser, M., & Mel, B. (2003). Dendrites: Bug or feature? Current Opinion in Neurobiology, 13, 372–383. Hill, D. N., Varga, Z., Jia, H., Sakmann, B., & Konnerth, A. (2013). Multibranch activity in basal and tuft dendrites during firing of layer 5 cortical neurons in vivo. Proceedings of the National Academy of Sciences of the United States of America, 110, 13618–13623. Hughes, E. G., Peng, X., Gleichman, A. J., Lai, M., Zhou, L., Tsou, R., ... Balice-Gordon, R. J. (2010). Cellular and synaptic mechanisms of anti-NMDAR encephalitis. Journal of Neuroscience, 30, 5866–5875. Jadi, J. P., Behabadi, B. F., Poleg-Polsky, A., Schiller, J., & Mel, B. W. (2014). An augmented two-layer model captures nonlinear analog spatial integration effects in pyramidal neuron dendrites. Proceedings of the IEEE, 102, 782–798. Jiang, X., Wang, G., Lee, A. J., Stornetta, R. L., & Zhu, J. J. (2013). The organization of two new cortical interneuronal circuits. Nature Neuroscience, 16, 210–218. Johnson, M. H. (2011). Interactive specialization: A domain-general framework for human functional brain development? Developmental Cognitive Neuroscience, 1, 7–21. Kay, J., Floreano, D., & Phillips, W. A. (1998). Contextually guided unsupervised learning using local multivariate binary processors. Neural Networks, 11, 117–140. Kay, J. (2000). Neural networks for unsupervised learning based on information theory. In J. W. Kay & D. M. Titterington (Eds.), Statistics and neural networks: Advances at the interface (pp. 25–63). Oxford: Oxford University Press. Kay, J., & Phillips, W. A. (1997). Activation functions, computational goals, and learning rules for local processors with contextual guidance. Neural Computation, 9, 895–910. Kay, J., & Phillips, W. A. (2011). Coherent infomax as a computational goal for neural systems. Bulletin of Mathematical Biology, 73, 344–372. http://dx.doi.org/ 10.1007/s11583-010-9564-x. Koch, C., & Segev, I. (2000). The role of single neurons in information processing. Nature Neuroscience, 3, 1171–1177. Körding, K. P., & König, P. (2000). Learning with two sites of synaptic integration. Network: Computation in Neural Systems, 11, 1–15. Kotchoubey, B., & El-Khoury, S. (2014). Event-related potentials indicate context effect in reading ambiguous words. Brain and Cognition, 92, 48–60. Kveraga, K., Ghuman, A. S., & Bar, M. (2007). Top-down predictions in the cognitive brain. Brain and Cognition, 65, 145–168. Lamme, V. A. F. (2004). Beyond the classical receptive field: Contextual modulation of V1 responses. In J. S. Werner & L. M. Chalupa (Eds.), The visual neurosciences (pp. 720–732). Cambridge, MA: MIT Press. Larkum, M. (2013). A cellular mechanism for cortical associations: An organizing principle for the cerebral cortex. Trends in Neurosciences, 36, 141–151.

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

14

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx

Larkum, M. E. (2013). The yin and yang of cortical layer 1. Trends in Neurosciences, 16, 114–115. Larkum, M. E., Nevian, T., Sandler, M., Polsky, A., & Schiller, J. (2009). Synaptic integration in tuft dendrites of layer 5 pyramidal neurons: A new unifying principle. Science, 325, 756–760. Larkum, M. E., Senn, W., & Lüscher, H.-R. (2004). Top-down dendritic input increases the gain of layer 5 pyramidal neurons. Cerebral Cortex, 14, 1059–1070;. http://dx.doi.org/10.1093/cercor/bhh065. Larkum, M. E., Waters, J., Sakmann, B., & Helmchen, F. (2007). Dendritic spikes in apical dendrites of neocortical layer 2/3 pyramidal neurons. Journal of Neuroscience, 27, 8999–9008. Larkum, M. E., Zhu, J. J., & Sakmann, B. (1999). A new cellular mechanism for coupling inputs arriving at different cortical layers. Nature, 98(6725), 338–341. Lawson, R. P., Rees, G., & Friston, K. J. (2014). An aberrant precision account of autism. Frontiers Human Neuroscience, 8. http://dx.doi.org/10.3389/fnhum. 2014.00302. Ledergerber, D., & Larkum, M. E. (2010). Properties of layer 6 pyramidal neuron apical dendrites. Journal of Neuroscience, 30(39), 13031–13044. Ledergerber, D., & Larkum, M. E. (2012). The time window for generation of dendritic pikes by coincidence of action potentials and EPSPs is layer specific in somatosensory cortex. PLoS One, 7(3), e33146. http://dx.doi.org/10.1371/ journal.pone.0033146. Lee, A. J., Wang, G., Jiang, X., Johnson, S. M., Hoang, E. T., Lanté, F., ... Zhu, J. J. (2014). Canonical organization of layer 1 neuron-led cortical inhibitory and disinhibitory interneuronal circuits. Cerebral Cortex. http://dx.doi.org/10.1093/ cercor/bhu020. Linsker, R. (1988). Self-organization in a perceptual network. Computer, 21, 105–107. Magee, J. C., & Cook, E. P. (2000). Somatic EPSP amplitude is independent of synapse location in hippocampal pyramidal neurons. Nature Neuroscience, 3, 895–903. Major, G., Larkum, M. E., & Schiller, J. (2013). Active properties of neocortical pyramidal neuron dendrites. Annual Review of Neuroscience, 36, 1–24. Markov, N. T., & Kennedy, H. (2013). The importance of being hierarchical. Current Opinion in Neurobiology, 23, 187–194. Markov, N. T., Vezoli, J., Chameau, P., Falchier, A., Quilodran, R., Huissoud, C., ... Kennedy, H. (2014). Anatomy of hierarchy: Feedforward and feedback pathways in macaque visual cortex. Journal of Comparative Neurology, 522, 225–259. http://dx.doi.org/10.1002/cne.23458. Markram, K., & Markram, H. (2010). The intense world theory – A unifying theory of the neurobiology of autism. Frontiers in Human Neuroscience, 4. http://dx.doi. org/10.3389/fnhum.2010.00224. McCulloch, W. S., & Pitts, W. (1943). A logical calculus of the ideas immanent in nervous activity. The Bulletin of Mathematical Biophysics, 5, 115–133. Mehta, A. D., Ulbert, I., & Schroeder, C. E. (2000). Intermodal selective attention in monkeys. I: Distribution and timing of effects across visual areas. Cerebral Cortex, 10, 343–358. Meredith, R. M., Holmgren, C. D., Weidum, M., Burnashev, N., & Mansvelder, H. D. (2007). Increased threshold for spike-timing-dependent plasticity is caused by unreliable calcium signaling in mice lacking fragile X gene FMR1. Neuron, 54, 627–638. Meyer, K. (2015). The role of dendritic signaling in the anesthetic suppression of consciousness. Anesthesiology, 122, 1415–1431. Moghaddam, B., & Javitt, D. (2012). From revolution to evolution: The glutamate hypothesis of schizophrenia and its implication for treatment. Neuropsychopharmacology Reviews, 37, 4–15. Moldakarimov, S., Bazhenov, M., & Sejnowski, T. J. (2010). Perceptual priming leads to reduction of gamma frequency oscillations. Proceedings of the National Academy of Sciences of the United States of America, 107, 5640–5645. http://dx. doi.org/10.1073/pnas.0907525107. Morey, C. C., & Bieler, M. (2013). Visual short-term memory always requires general attention. Psychonomic Bulletin & Review. http://dx.doi.org/10.3758/s13423012-0313-z. Muckli, L., & Petro, L. S. (2013). Network interactions: Non-geniculate input to V1. Current Opinion in Neurobiology, 23, 195–201. Murayama, M., & Larkum, M. E. (2009). Enhanced dendritic activity in awake rats. Proceedings of the National Academy of Sciences of the United States of America, 106, 20482–20486. http://dx.doi.org/10.1073/pnas.0910379106. Murayama, M., Pérez-Garci, E., Nevian, T., Bock, T., Senn, W., & Larkum, M. E. (2009). Dendritic encoding of sensory stimuli controlled by deep cortical interneurons. Nature, 457(7233), 1137–1141. Ochsner, K. N., & Gross, J. J. (2005). The cognitive control of emotion. Trends in Cognitive Sciences, 9, 242–249. Ondobaka, S., Kilner, J., & Friston, K. (2015). The role of interoceptive inference in theory of mind (this issue). Palmer, S. E. (1975). The effects of contextual scenes on the identification of objects. Memory & Cognition, 3, 519–526. Palmer, L. M., Schulz, J. M., Murphy, S. C., Ledergerber, D., Murayama, M., & Larkum, M. E. (2012). The cellular basis of GABAB-mediated interhemispheric inhibition. Science, 335, 989–993. Palmer, L. M., Shai, A. S., Reeve, J. E., Andersen, H. L., Paulsen, O., & Larkum, M. E. (2014). NMDA spikes enhance action potential generation during sensory input. Nature Neuroscience, 17, 383–390. Pérez-Garci, E., Gassmann, M., Bettler, B., & Larkum, M. E. (2006). The GABA B1b isoform mediates long-lasting inhibition of dendritic Ca2+ spikes in layer 5 somatosensory pyramidal neurons. Neuron, 50, 603–616.

Pessoa, L. (2008). On the relationship between emotion and cognition. Nature Reviews Neuroscience, 9, 148–158. Petreanu, L., Mao, T., Sternson, S. M., & Svoboda, K. (2009). The subcellular organization of neocortical excitatory connections. Nature, 457, 1142–1145. Phillips, W. A. (1983). Short-term visual memory. Philosophical transactions of the Royal Society of London, B, 302, 295–309. Phillips, W. A. (2012). Self-organized complexity and Coherent Infomax from the viewpoint of Jaynes’s probability theory. Information, 3, 1–15. http://dx.doi.org/ 10.3390/info3010001. Phillips, W. A. (2015). Prenatal exposure to alcohol causes enduring brain damage. Adoption & Fostering, 39(3) (in press). Phillips, W. A., & Christie, D. F. (1977). Interference with visualization. The Quarterly Journal of Experimental Psychology, 29, 637–650. Phillips, W. A., Clark, A., & Silverstein, S. M. (2015). On the functions, mechanisms, and malfunctions of intracortical contextual modulation. Neuroscience and Biobehavioral Reviews, 52, 1–20. http://dx.doi.org/10.1016/j.neubiorev. 2015.02.010. Phillips, W. A., & Craven, B. (2000). Interactions between coincident and orthogonal cues to texture boundaries. Perception & Psychophysics, 62, 1019–1038. Phillips, W. A., Kay, J., & Smyth, D. (1995). The discovery of structure by multi-stream networks of local processors with contextual guidance. Network: Computation in Neural Systems, 6, 225–246. Phillips, W. A., & Silverstein, S. M. (2003). Convergence of biological and psychological perspectives on cognitive coordination in schizophrenia. Behavioral and Brain Sciences, 26, 65–138. Phillips, W. A., & Silverstein, S. M. (2013). The coherent organization of mental life depends on mechanisms for context-sensitive gain-control that are impaired in schizophrenia. Frontiers in Psychology, 4, 307. http://dx.doi.org/10.3389/ fpsyg.2013.00307. Phillips, W. A., & Singer, W. (1997). In search of common foundations for cortical computation. Behavioral and Brain Sciences, 20, 657–722. Phillips, W. A., von der Malsburg, C., & Singer, W. (2010). Dynamic coordination in brain and mind. In C. von der Malsburg, W. A. Phillips, & W. Singer (Eds.). Dynamic coordination in the brain: From neurons to mind (Strüngmann forum report: Vol. 5, pp. 1–24). Cambridge, MA: MIT Press. Pike, F. G., Meredith, R. M., Olding, A. W. A., & Paulsen, O. (1999). Postsynaptic bursting is essential for ‘Hebbian’ induction of associative long-term potentiation at excitatory synapses in rat hippocampus. The Journal of Physiology (London), 518, 571–576. Potez, S., & Larkum, M. E. (2008). Effect of common anesthetics on dendritic properties in layer 5 neocortical pyramidal neurons. Journal of Neurophysiology, 99, 1394–1407. http://dx.doi.org/10.1152/jn.01126.2007. Qi, G., & Feldmeyer, D. (2015). Dendritic target region-specific formation of synapses between excitatory layer 4 neurons and layer 6 pyramidal cells. Cerebral Cortex. http://dx.doi.org/10.1093/cercor/bhu334. Reynolds, J. H., & Heeger, D. J. (2009). The normalization model of attention. Neuron, 61, 168–185. Roth, J. K., Serences, J. T., & Courtney, S. M. (2006). Neural system for controlling the contents of object working memory in humans. Cerebral Cortex, 16, 1595–1603. Rothenstein, A. L., & Tsotsos, J. K. (2014). Attentional modulation and selection – An integrated approach. PLoS One, 9(6), e99681. Rubio-Garrido, P., Perez-de-Manzo, F., Porrero, C., Galazo, M. J., & Clasca, F. (2009). Thalamic input to distal apical dendrites in neocortical layer 1 is massive and highly convergent. Cerebral Cortex, 19, 2380–2395. Ryan, T. J., Kopanitsa, M. V., Indersmitten, T., Nithianantharajah, J., Afinowi, N. O., Pettit, C., ... Komiyama, N. H. (2013). Evolution of GluN2A/B cytoplasmic domains diversified vertebrate synaptic plasticity and behavior. Nature Neuroscience, 16, 25–32. Salinas, E., & Sejnowski, T. J. (2001). Gain modulation in the central nervous system: Where behavior, neurophysiology, and computation meet. Neuroscientist, 7, 430–440. Salinas, E., & Their, P. (2000). Gain modulation: A major computational principle of the central nervous system. Neuron, 27, 15–21. Schiller, J., Major, G., Koester, H. J., & Schiller, Y. (2000). NMDA spikes in basal dendrites of cortical pyramidal neurons. Nature, 404, 285–289. Self, M. W., Kooijmansa, R. N., Supèr, H., Lamme, V. A., & Roelfsema, P. R. (2012). Different glutamate receptors convey feedforward and recurrent processing in macaque V1. Proceedings of the National Academy of Sciences of the United States of America, 109, 11031–11036. http://dx.doi.org/10.1073/pnas.1119527109. Seth, A. K. (2013). Interoceptive inference, emotion, and the embodied self. Trends in Cognitive Sciences, 17, 565–573. Seth, A. K., Suzuki, K., & Critchley, H. D. (2011). An interoceptive predictive coding model of conscious presence. Frontiers in Psychology, 2, 395. Shah, A. S., Bressler, S. L., Knuth, K. H., Ding, M., Mehta, A. D., Ulbert, I., & Schroeder, C. E. (2004). Neural dynamics and the fundamental mechanisms of eventrelated brain potentials. Cerebral Cortex, 14, 476–483. Shao, Z. W., & Burkhalter, A. (1996). Different balance of excitation and inhibition in forward and feedback circuits of rat visual cortex. Journal of Neuroscience, 16, 7353–7365. Shepherd, G. M. (2011). The microcircuit concept applied to cortical evolution: From three-layer to six-layer cortex. Frontiers in Neuroanatomy, 5, 30. http://dx. doi.org/10.3389/fnana.2011.00030. Sherman, S. M. (2012). Thalamocortical interactions. Current Opinion in Neurobiology, 22(4), 575–579. Shipp, S., Adams, R. A., & Friston, K. J. (2013). Reflections on agranular architecture: Predictive coding in the motor cortex. Trends in Neurosciences, 36, 706–716.

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005

W.A. Phillips / Brain and Cognition xxx (2015) xxx–xxx Shipp, S., Adams, D. L., Moutoussis, K., & Zeki, S. (2009). Feature binding in the feedback layers of area V2. Cerebral Cortex, 19, 2230–2239. http://dx.doi.org/ 10.1093/cercor/bhn243. Shlosberg, D., Amitai, Y., & Azouz, R. (2006). Time-dependent, layer-specific modulation of sensory responses mediated by neocortical layer 1. Journal of Neurophysiology, 96, 3170–3182. Siegel, M., Körding, K. P., & König, P. (2000). Integrating top-down and bottom-up sensory processing by somato-dendritic interactions. Journal of Computational Neuroscience, 8, 161–173. Smyth, D., Phillips, W. A., & Kay, J. (1996). Measures for investigating the contextual modulation of information transmission. Network: Computation in Neural Systems, 7, 307–316. Sohn, W., Papathomas, T. V., Blaser, E., & Vidnyanszky, Z. (2004). Object-based crossfeature attentional modulation from color to motion. Vision Research, 44, 1437–1443. Spratling, M. W. (2002). Cortical region interactions and the functional role of apical dendrites. Behavioral and Cognitive Neuroscience Reviews, 1, 219–228. Spratling, M. W. (2008). Predictive-coding as a model of biased competition in visual attention. Vision Research, 48, 1391–1408. Spratling, M. W. (2010). Predictive coding as a model of response properties in cortical area V1. Journal of Neuroscience, 30, 3531–3543. Spratling, M. W. (2012). Unsupervised learning of generative and discriminative weights encoding elementary image components in a predictive coding model of cortical function. Neural Computation, 24, 60–103. Spratling, M. W. (2014). A single functional model of drivers and modulators in cortex. Journal of Computational Neuroscience, 36, 97–118. Spratling, M. W., & Johnson, M. H. (2004). A feedback model of visual attention. Journal of Cognitive Neuroscience, 16, 219–237. Spratling, M. W., & Johnson, M. H. (2006). A feedback model of perceptual learning and categorisation. Visual Cognition, 13, 129–165. Troscianko, T., Davidoff, J., Humphreys, G., Landis, T., Fahle, M., Greenlee, M., ... Phillips, W. A. (1996). Human colour discrimination based on a nonparvocellular pathway. Current Biology, 6, 200–210. Ullman, S. (1995). Sequence seeking and counter streams: A computational model for bidirectional information flow in the visual cortex. Cerebral Cortex, 5(1), 1–11.

15

van Kerkoerle, T., Self, M. W., Dagnino, B., Gariel-Mathis, M. A., Poort, J., van der Togt, C., & Roelfsema, P. R. (2014). Alpha and gamma oscillations characterize feedback and feedforward processing in monkey visual cortex. Proceedings of the National Academy of Sciences of the United States of America, 111, 14332–14341. von der Malsburg, C., Phillips, W. A., & Singer, W. (Eds.). (2010). Dynamic coordination in the brain: From neurons to mind (Strüngmann forum report: Vol. 5). Cambridge, MA: MIT Press. Wang, Y., Toledo-Rodriguez, M., Gupta, A., Wu, C., Silberberg, G., Luo, J., & Markram, H. (2004). Anatomical, physiological and molecular properties of Martinotti cells in the somatosensory cortex of the juvenile rat. Journal of Physiology, 561, 65–90. http://dx.doi.org/10.1113/jphysiol.2004.073353. Waters, J., Larkum, M., Sakmann, B., & Helmchen, F. (2003). Supralinear Ca2+ influx into dendritic tufts of layer 2/3 neocortical pyramidal neurons in vitro and in vivo. Journal of Neuroscience, 23, 8558–8567. Wibral, M., Priesemann, V., Kay, J. W., Lizier, J. T., & Phillips, W. A. (2015). Partial information decomposition as a unified approach to the specification of neural goal functions (this issue). Wiggs, C. L., & Martin, A. (1998). Properties and mechanisms of perceptual priming. Current Opinion in Neurobiology, 8, 227–233. Williams, S. R., & Stuart, G. J. (2003). Role of dendritic synapse location in the control of action potential output. Trends in Neurosciences, 26, 147–154. Xu, N.-L., Harnett, M. T., Williams, S. R., Huber, D., O’Connor, D. H., Svobda, K., & Magee, J. C. (2012). Nonlinear dendritic integration of sensory and motor input during an active sensing task. Nature, 492(7428), 247–251. Zanto, T. P., Rubens, M. T., Thangavel, A., & Gazzaley, A. (2011). Causal role of the prefrontal cortex in top-down modulation of visual processing and working memory. Nature Neuroscience, 14, 656–661. Zhang, Y., Bonnan, A., Bony, G., Ferezou, I., Pietropaolo, S., Ginger, M., ... Frick, A. (2014). Dendritic channelopathies contribute to neocortical and sensory hyperexcitability in Fmr1-/y mice. Nature Neuroscience, 17, 1701–1709. Zipser, K., Lamme, V. A. F., & Schiller, P. H. (1996). Contextual modulation in primary visual cortex. Journal of Neuroscience, 15, 7376–7389.

Please cite this article in press as: Phillips, W. A. Cognitive functions of intracellular mechanisms for contextual amplification. Brain and Cognition (2015), http://dx.doi.org/10.1016/j.bandc.2015.09.005