- Email: [email protected]
How Do We Keep Information ‘Online’?
(Figure 1C). Specifically, a region's loading on the principal gradient is strongly predicted by its distance to peak locations within the default network. In other words, regions that tend to be maximally dissimilar from these higher-order association zones in terms of their connectivity also tend to lie furthest from them spatially. The component explaining the second-most variance distinguishes sensory modalities, with the primary visual cortex at one end and the somatosensory and motor cortices at the other. The authors find a comparable pattern when repeating the analysis on macaque tract-tracing data: the first component similarly groups sensory areas together on one end of the gradient and puts anterior cingulate, prefrontal, and temporal regions on the other (in the macaque data, the posterior cingulate does not group with the other association regions, but that may relate to the resolution of the database). They also use an independent fMRI task meta-analysis resource (NeuroSynth) to provide further evidence that a region's score along the principal component is related to its functional characterization in humans. Terms associated with sensory processing and movement load on regions positioned at one end of the first axis; terms associated with autobiographical memory, social cognition, emotion, and language load on regions at the other end of the axis.
abrupt. The results of Margulies et al. suggest that there is a macrolevel organization of connectivity patterns that in some cases might be better described as an array of gradients that move smoothly across the cortical surface, anchored by primary sensory and motor regions. An exciting avenue for further exploration will be to examine these and other properties of ‘macroscale’ or ‘superareal’ spatial organization in other species. A putative default network has been observed in other mammals besides primates [7] and recent work suggests that some principles that shape connectivity in rodents and primates are strikingly conserved [8]. Another direction will be to better understand the nature of interactions within and across regions that occupy adjacent or distal positions along the principal gradient. The approach taken by Margulies et al. does not distinguish between regions that are highly functionally coupled because of many connections or because of a few very strong connections, but these two forms of network architecture may have different implications for how information is represented or propagated through local and distributed circuits. Overall, the model described by Margulies et al. offers a simple and powerful explanation of one constraint that structures the functional topology of the primate neocortex. This work reinforces the need for neurodevelopmental accounts that specify mechanisms for how regions at the far end of the gradient, including the default network, come to be connected to one another across such vast distances.
With some exceptions methods for describing the topology of connectivity patterns across the brain have largely approached the problem by attempting to parcellate or subdivide functional systems. Much as the earlier work by neuroanatomists partitioned the neocortex into Acknowledgments discrete areas with sharp boundaries The authors would like to thank Marta Florio for helpful defined by cytoarchitecture, efforts that discussions. identify discrete network boundaries 1 Department of Genetics, Harvard Medical School, with connectivity data make it easier to Boston, MA, USA categorize, name, and apply parcella- 2Broad Institute, Cambridge, MA, USA 3 tions [5,6]. However, this can set up Department of Anthropology, Center for the Advanced Study of Human Paleobiology and GW Mind–Brain the expectation that transitions between Institute, The George Washington University, Washington, connectivity networks are sharp and DC, USA
*Correspondence: [email protected] (F.M. Krienen) and [email protected] (C.C. Sherwood). http://dx.doi.org/10.1016/j.tics.2016.12.002 References 1. Kaas, J.H. (2013) The evolution of brains from early mammals to humans. Wiley Interdiscip. Rev. Cogn. Sci. 4, 33–45 2. Krubitzer, L. (2009) In search of a unifying theory of complex brain evolution. Ann. N. Y. Acad. Sci. 1156, 44–67 3. Margulies, D.S. et al. (2016) Situating the default-mode network along a principal gradient of macroscale cortical organization. Proc. Natl. Acad. Sci. U.S.A. 113, 12574– 12579 4. Rosa, M.G.P. (2002) Visual maps in the adult primate cerebral cortex: some implications for brain development and evolution. Braz. J. Med. Biol. Res. 35, 1485–1498 5. Yeo, B.T.T. et al. (2011) The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J. Neurophysiol. 106, 1125–1165 6. Glasser, M.F. et al. (2016) A multi-modal parcellation of human cerebral cortex. Nature 536, 171–178 7. Stafford, J.M. et al. (2014) Large-scale topology and the default mode network in the mouse connectome. Proc. Natl. Acad. Sci. U.S.A. 111, 18745–18750 8. Horvát, S. et al. (2016) Spatial embedding and wiring cost constrain the functional layout of the cortical network of rodents and primates. PLoS Biol. 14, e1002512 9. Krienen, F.M. and Buckner, R.L. Human association cortex: expanded, untethered, neotenous, and plastic. In Evolution of Nervous Systems (2nd edn) (Kaas, J., ed.), pp. 1-15, Elsevier.(in press).
Spotlight
How Do We Keep Information ‘Online’? David Soto1,2,*,@ New magnetoencephalography (MEG) results indicate that a putative marker of conscious processes – namely, the global broadcasting of information across large-scale cortical networks – can also operate during the maintenance of nonconscious input. I discuss the implications for the theoretical linkage between conscious awareness and working memory functions. A fundamental problem in psychology and neuroscience is understanding how the contents of the mind develop on a moment-to-moment basis. Intuitively, awareness seems a prerequisite to keep relevant information online and guide
Trends in Cognitive Sciences, February 2017, Vol. 21, No. 2
63
King and colleagues [2] put these models to test by applying machine learning techniques to MEG data recorded while human participants were tasked to remember the orientation of a brief and masked stimulus during a short maintenance period (Figure 1A). Even when observers reported no experience of (A)
Target
Mask
seeing the stimulus, memory performance was significantly above chance (Figure 1B), in keeping with the view that non-conscious processes may pervade executive control and higher-order mnemonic functions [3,4]. Using MEG data from single trials, a decoder was trained to predict features of the memory target (i. e., the physical absence versus presence of the item, its contrast, orientation). Even when observers reported no experience of the stimulus, its presence/absence could be decoded during the maintenance period. Most notably, a decoder trained to discriminate stimulus presence/ absence during the early stages of maintenance generalized across later processing stages equally well in both conscious and non-conscious trials (Figure 1C). These temporal generalization results also showed that the later maintenance stages
Probe
Delay
Tilt?
8
50 ms
800 ms
?
8
117 ms
A ‘threshold’ model of consciousness would posit that non-conscious processes are a reflection of conscious counterparts but mainly differ in magnitude (e.g., similar neural processing substrates/hierarchies but attenuated activity in non-conscious states). In keeping with this ‘threshold’ view, decoding of the presence/absence of the memory target during maintenance was greater with higher visibility ratings (Figure 1D). Also, the temporal generalization results were suggestive of similar
Visibility?
? 17 ms
are increasingly metastable and encode the presence of the target for much longer durations than the early stages. This is consonant with a neural architecture in which non-conscious memoranda spread and feed back across multiple processing stages in a similar way to their conscious counterparts.
(B)
100%
∗
Accuracy
behavior. Accordingly, influential neural models propose that information is selected and kept available for conscious inspection and higher-order cognitive processes by virtue of sustained activity in parietofrontal networks (cf. global neuronal workspace theory) and recurrent neural feedback loops [1]. It is often implied that non-conscious input may trigger only feedforward neural responses that typically fail to support highorder cognitive processes requiring online maintenance of information.
∗ ∗
∗ 50%
(C)
.11
(D)
.89 AUC
Target presence Visibility 3 Visibility 0
Visibility Baseline
Early
0 1 2 3
Delay
Probe
Target
∗
AUC
∗
Target presence
∗ ∗∗∗
∗∗
∗∗
∗∗
∗∗
0.12 Target angle Probe Tesng me
rad.
Training me (ms)
1.0
∗ ∗∗ ∗
∗∗ ∗ ∗ ∗ ∗
∗
100 ms
Figure 1. Illustration of the Experimental Protocol and Results. (A) Example of a trial in the task. On target-present trials, oriented gratings (the memory target) appeared for 17 ms, masked, and were followed by a test grating 800 ms later. Participants had to report whether the test was oriented clockwise or anticlockwise relative to the memory target and the visibility of the masked stimulus on each trial (‘0’ means no experience of stimulus). (B) Behavioral results. The bar graph shows the proportion of correct responses as a function of visibility. (C) Results from the temporal generalization analyses, in which a decoder of the presence/absence of the target is fitted from magnetoencephalography (MEG) data at a given timeframe and tested in another timeframe. Classification performance is measured by the area under the receiver operating characteristic (ROC) curve (AUC). (D) Decoding estimates of target presence (top) and orientation (bottom) across each time period of the trial and visibility ratings, with significant levels marked by *.
64
Trends in Cognitive Sciences, February 2017, Vol. 21, No. 2
processing architectures for non-conscious and conscious maintenance; however, the generalization of the decoder trained in the later stages of maintenance was lower and of shorter duration on nonconscious trials (see Figure 1C). This begs the fundamental question of whether distinct neural coding schemes support non-conscious and conscious maintenance. An intriguing result was that decoding of the relevant memory feature (i.e., orientation) on non-conscious trials was highly reliable following the stimulus but far weaker during the maintenance period; however, it recurred again with the onset of the test (Figure 1D, bottom right). Obviously the orientation of the nonconscious item was somehow retained in the brain, as participants used this information to support above-chance performance in the delayed discrimination test. However, and contrary to the conscious trials, there was little evidence that nonconscious trials were characterized by a sustained neural response that coded the relevant orientation feature (although the error in decoding was biased towards the orientation of the target during the maintenance phase). One possibility is that activity-silent neural coding schemes [5] (e.g., via modulation of synaptic weights in the absence of persistent neuronal firing) are most characteristic of information maintenance processes that are decoupled from conscious awareness. That decoding of the relevant orientation feature tended to wax and wane during non-conscious maintenance in King et al.’s study may not, therefore, be due to low signal to noise but could relate to a distinct neural coding scheme relative to conscious maintenance. This result also argues against the view that ‘non-conscious’ maintenance is mediated by observers making a conscious guess about the non-conscious stimulus and then keeping it online [6], as then orientation decoding should have been greater than chance
level during the maintenance phase of across visual and verbal domains of the non-conscious trials. increasing complexity and assessing whether the maintained information can Relevant clues to disassociate non-con- be further manipulated in a goal-directed scious versus conscious maintenance manner. The approach of King et al. cermechanisms could come from brain local- tainly paves the way for the tracking of ization analyses of the MEG signals, which moment-to-moment changes in mental in King et al.’s study involved a network of representations across states of (un)awarevisual areas following the stimulus, ness, but tackling the above issues in future extending to the parietal cortex during work will boost the ramifications of this maintenance and frontal substrates during approach for our understanding of the decision-making; however, given the bases of working memory and higher-order known difficulties in localizing brain sour- mnemonic functions [10] and their critical ces in MEG, defining the anatomy of con- role in guiding behavior. scious and non-conscious maintenance awaits further investigation. Future work Acknowledgments could exploit recent developments in merg- D.S. acknowledges financial support from the Spanish ing neural representation similarity analyses in Ministry of Economy and Competitiveness through the MEG and fMRI data [7]. Assessing the neural ‘Severo Ochoa’ Programme for Centres/Units of representational spaces of conscious and Excellence in R&D (SEV-2015-490). non-conscious contents across longer main- 1Basque Center on Cognition, Brain, and Language, San tenance periods than those used by King Sebastian, Spain 2 Ikerbasque, Basque Foundation for Science, Bilbao, et al. would help to define the spatiotemporal Spain neural dynamics of maintenance across *Correspondence: [email protected] (D. Soto). states of (un)awareness. @
It is also important not to underestimate the multiplicity of contexts that are relevant to the study of the interplay between working memory and conscious awareness. Memory paradigms using visible items demonstrated how memoranda may adopt different states of access for goal-directed behavior depending on attention [8]. An outstanding issue is whether this is extensible to the case of non-conscious maintenance and how this is modulated by the observer's attention state during the encoding of the relevant information or the withdrawal of attention during the maintenance period, since the single memory item in King et al.’s study always appeared at an attended location. Also, the limits and scope of brain maintenance mechanisms remain to be tested with observers who have objectively null sensitivity to the presence/ identity of the information [9]. More broadly, additional experiments are needed to pinpoint the domain generality/specificity of the underlying mechanisms; for instance, testing non-conscious neural maintenance
Twitter: @d_soto_b
http://dx.doi.org/10.1016/j.tics.2016.12.009 References 1. Dehaene, S. and Changeux, J.P. (2011) Experimental and theoretical approaches to conscious processing. Neuron 70, 200–227 2. King, J.-R. et al. (2016) Brain mechanisms underlying the brief maintenance of seen and unseen sensory information. Neuron 92, 1122–1134 3. Soto, D. and Silvanto, J. (2014) Reappraising the relationship between working memory and conscious awareness. Trends Cogn. Sci. 18, 520–525 4. Rosenthal, C.R. and Soto, D. (2016) The anatomy of nonconscious recognition memory. Trends Neurosci. 39, 707–711 5. Stokes, M.G. (2015) ‘Activity-silent’ working memory in prefrontal cortex: a dynamic coding framework. Trends Cogn. Sci. 19, 394–405 6. Stein, T. et al. (2016) Can working memory be non-conscious? Neurosci. Conscious. 2016, niv011 7. Cichy, R.M. et al. (2016) Similarity-based fusion of MEG and fMRI reveals spatio-temporal dynamics in human cortex during visual object recognition. Cereb. Cortex 26, 3563–3579 8. LaRocque, J.J. et al. (2014) Multiple neural states of representation in short-term memory? It's a matter of attention. Front. Hum. Neurosci. 8, 5 9. Pan, Y. et al. (2014) Working memory biasing of visual perception without awareness. Atten. Percept. Psychophys. 76, 2051–2062 10. Baddeley, A. (1993) Working memory and conscious awareness. In Theories of Memory, pp. 11–28, Lawrence Erlbaum Associates
Trends in Cognitive Sciences, February 2017, Vol. 21, No. 2
65
abrupt. The results of Margulies et al. suggest that there is a macrolevel organization of connectivity patterns that in some cases might be better described as an array of gradients that move smoothly across the cortical surface, anchored by primary sensory and motor regions. An exciting avenue for further exploration will be to examine these and other properties of ‘macroscale’ or ‘superareal’ spatial organization in other species. A putative default network has been observed in other mammals besides primates [7] and recent work suggests that some principles that shape connectivity in rodents and primates are strikingly conserved [8]. Another direction will be to better understand the nature of interactions within and across regions that occupy adjacent or distal positions along the principal gradient. The approach taken by Margulies et al. does not distinguish between regions that are highly functionally coupled because of many connections or because of a few very strong connections, but these two forms of network architecture may have different implications for how information is represented or propagated through local and distributed circuits. Overall, the model described by Margulies et al. offers a simple and powerful explanation of one constraint that structures the functional topology of the primate neocortex. This work reinforces the need for neurodevelopmental accounts that specify mechanisms for how regions at the far end of the gradient, including the default network, come to be connected to one another across such vast distances.
With some exceptions methods for describing the topology of connectivity patterns across the brain have largely approached the problem by attempting to parcellate or subdivide functional systems. Much as the earlier work by neuroanatomists partitioned the neocortex into Acknowledgments discrete areas with sharp boundaries The authors would like to thank Marta Florio for helpful defined by cytoarchitecture, efforts that discussions. identify discrete network boundaries 1 Department of Genetics, Harvard Medical School, with connectivity data make it easier to Boston, MA, USA categorize, name, and apply parcella- 2Broad Institute, Cambridge, MA, USA 3 tions [5,6]. However, this can set up Department of Anthropology, Center for the Advanced Study of Human Paleobiology and GW Mind–Brain the expectation that transitions between Institute, The George Washington University, Washington, connectivity networks are sharp and DC, USA
*Correspondence: [email protected] (F.M. Krienen) and [email protected] (C.C. Sherwood). http://dx.doi.org/10.1016/j.tics.2016.12.002 References 1. Kaas, J.H. (2013) The evolution of brains from early mammals to humans. Wiley Interdiscip. Rev. Cogn. Sci. 4, 33–45 2. Krubitzer, L. (2009) In search of a unifying theory of complex brain evolution. Ann. N. Y. Acad. Sci. 1156, 44–67 3. Margulies, D.S. et al. (2016) Situating the default-mode network along a principal gradient of macroscale cortical organization. Proc. Natl. Acad. Sci. U.S.A. 113, 12574– 12579 4. Rosa, M.G.P. (2002) Visual maps in the adult primate cerebral cortex: some implications for brain development and evolution. Braz. J. Med. Biol. Res. 35, 1485–1498 5. Yeo, B.T.T. et al. (2011) The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J. Neurophysiol. 106, 1125–1165 6. Glasser, M.F. et al. (2016) A multi-modal parcellation of human cerebral cortex. Nature 536, 171–178 7. Stafford, J.M. et al. (2014) Large-scale topology and the default mode network in the mouse connectome. Proc. Natl. Acad. Sci. U.S.A. 111, 18745–18750 8. Horvát, S. et al. (2016) Spatial embedding and wiring cost constrain the functional layout of the cortical network of rodents and primates. PLoS Biol. 14, e1002512 9. Krienen, F.M. and Buckner, R.L. Human association cortex: expanded, untethered, neotenous, and plastic. In Evolution of Nervous Systems (2nd edn) (Kaas, J., ed.), pp. 1-15, Elsevier.(in press).
Spotlight
How Do We Keep Information ‘Online’? David Soto1,2,*,@ New magnetoencephalography (MEG) results indicate that a putative marker of conscious processes – namely, the global broadcasting of information across large-scale cortical networks – can also operate during the maintenance of nonconscious input. I discuss the implications for the theoretical linkage between conscious awareness and working memory functions. A fundamental problem in psychology and neuroscience is understanding how the contents of the mind develop on a moment-to-moment basis. Intuitively, awareness seems a prerequisite to keep relevant information online and guide
Trends in Cognitive Sciences, February 2017, Vol. 21, No. 2
63
King and colleagues [2] put these models to test by applying machine learning techniques to MEG data recorded while human participants were tasked to remember the orientation of a brief and masked stimulus during a short maintenance period (Figure 1A). Even when observers reported no experience of (A)
Target
Mask
seeing the stimulus, memory performance was significantly above chance (Figure 1B), in keeping with the view that non-conscious processes may pervade executive control and higher-order mnemonic functions [3,4]. Using MEG data from single trials, a decoder was trained to predict features of the memory target (i. e., the physical absence versus presence of the item, its contrast, orientation). Even when observers reported no experience of the stimulus, its presence/absence could be decoded during the maintenance period. Most notably, a decoder trained to discriminate stimulus presence/ absence during the early stages of maintenance generalized across later processing stages equally well in both conscious and non-conscious trials (Figure 1C). These temporal generalization results also showed that the later maintenance stages
Probe
Delay
Tilt?
8
50 ms
800 ms
?
8
117 ms
A ‘threshold’ model of consciousness would posit that non-conscious processes are a reflection of conscious counterparts but mainly differ in magnitude (e.g., similar neural processing substrates/hierarchies but attenuated activity in non-conscious states). In keeping with this ‘threshold’ view, decoding of the presence/absence of the memory target during maintenance was greater with higher visibility ratings (Figure 1D). Also, the temporal generalization results were suggestive of similar
Visibility?
? 17 ms
are increasingly metastable and encode the presence of the target for much longer durations than the early stages. This is consonant with a neural architecture in which non-conscious memoranda spread and feed back across multiple processing stages in a similar way to their conscious counterparts.
(B)
100%
∗
Accuracy
behavior. Accordingly, influential neural models propose that information is selected and kept available for conscious inspection and higher-order cognitive processes by virtue of sustained activity in parietofrontal networks (cf. global neuronal workspace theory) and recurrent neural feedback loops [1]. It is often implied that non-conscious input may trigger only feedforward neural responses that typically fail to support highorder cognitive processes requiring online maintenance of information.
∗ ∗
∗ 50%
(C)
.11
(D)
.89 AUC
Target presence Visibility 3 Visibility 0
Visibility Baseline
Early
0 1 2 3
Delay
Probe
Target
∗
AUC
∗
Target presence
∗ ∗∗∗
∗∗
∗∗
∗∗
∗∗
0.12 Target angle Probe Tesng me
rad.
Training me (ms)
1.0
∗ ∗∗ ∗
∗∗ ∗ ∗ ∗ ∗
∗
100 ms
Figure 1. Illustration of the Experimental Protocol and Results. (A) Example of a trial in the task. On target-present trials, oriented gratings (the memory target) appeared for 17 ms, masked, and were followed by a test grating 800 ms later. Participants had to report whether the test was oriented clockwise or anticlockwise relative to the memory target and the visibility of the masked stimulus on each trial (‘0’ means no experience of stimulus). (B) Behavioral results. The bar graph shows the proportion of correct responses as a function of visibility. (C) Results from the temporal generalization analyses, in which a decoder of the presence/absence of the target is fitted from magnetoencephalography (MEG) data at a given timeframe and tested in another timeframe. Classification performance is measured by the area under the receiver operating characteristic (ROC) curve (AUC). (D) Decoding estimates of target presence (top) and orientation (bottom) across each time period of the trial and visibility ratings, with significant levels marked by *.
64
Trends in Cognitive Sciences, February 2017, Vol. 21, No. 2
processing architectures for non-conscious and conscious maintenance; however, the generalization of the decoder trained in the later stages of maintenance was lower and of shorter duration on nonconscious trials (see Figure 1C). This begs the fundamental question of whether distinct neural coding schemes support non-conscious and conscious maintenance. An intriguing result was that decoding of the relevant memory feature (i.e., orientation) on non-conscious trials was highly reliable following the stimulus but far weaker during the maintenance period; however, it recurred again with the onset of the test (Figure 1D, bottom right). Obviously the orientation of the nonconscious item was somehow retained in the brain, as participants used this information to support above-chance performance in the delayed discrimination test. However, and contrary to the conscious trials, there was little evidence that nonconscious trials were characterized by a sustained neural response that coded the relevant orientation feature (although the error in decoding was biased towards the orientation of the target during the maintenance phase). One possibility is that activity-silent neural coding schemes [5] (e.g., via modulation of synaptic weights in the absence of persistent neuronal firing) are most characteristic of information maintenance processes that are decoupled from conscious awareness. That decoding of the relevant orientation feature tended to wax and wane during non-conscious maintenance in King et al.’s study may not, therefore, be due to low signal to noise but could relate to a distinct neural coding scheme relative to conscious maintenance. This result also argues against the view that ‘non-conscious’ maintenance is mediated by observers making a conscious guess about the non-conscious stimulus and then keeping it online [6], as then orientation decoding should have been greater than chance
level during the maintenance phase of across visual and verbal domains of the non-conscious trials. increasing complexity and assessing whether the maintained information can Relevant clues to disassociate non-con- be further manipulated in a goal-directed scious versus conscious maintenance manner. The approach of King et al. cermechanisms could come from brain local- tainly paves the way for the tracking of ization analyses of the MEG signals, which moment-to-moment changes in mental in King et al.’s study involved a network of representations across states of (un)awarevisual areas following the stimulus, ness, but tackling the above issues in future extending to the parietal cortex during work will boost the ramifications of this maintenance and frontal substrates during approach for our understanding of the decision-making; however, given the bases of working memory and higher-order known difficulties in localizing brain sour- mnemonic functions [10] and their critical ces in MEG, defining the anatomy of con- role in guiding behavior. scious and non-conscious maintenance awaits further investigation. Future work Acknowledgments could exploit recent developments in merg- D.S. acknowledges financial support from the Spanish ing neural representation similarity analyses in Ministry of Economy and Competitiveness through the MEG and fMRI data [7]. Assessing the neural ‘Severo Ochoa’ Programme for Centres/Units of representational spaces of conscious and Excellence in R&D (SEV-2015-490). non-conscious contents across longer main- 1Basque Center on Cognition, Brain, and Language, San tenance periods than those used by King Sebastian, Spain 2 Ikerbasque, Basque Foundation for Science, Bilbao, et al. would help to define the spatiotemporal Spain neural dynamics of maintenance across *Correspondence: [email protected] (D. Soto). states of (un)awareness. @
It is also important not to underestimate the multiplicity of contexts that are relevant to the study of the interplay between working memory and conscious awareness. Memory paradigms using visible items demonstrated how memoranda may adopt different states of access for goal-directed behavior depending on attention [8]. An outstanding issue is whether this is extensible to the case of non-conscious maintenance and how this is modulated by the observer's attention state during the encoding of the relevant information or the withdrawal of attention during the maintenance period, since the single memory item in King et al.’s study always appeared at an attended location. Also, the limits and scope of brain maintenance mechanisms remain to be tested with observers who have objectively null sensitivity to the presence/ identity of the information [9]. More broadly, additional experiments are needed to pinpoint the domain generality/specificity of the underlying mechanisms; for instance, testing non-conscious neural maintenance
Twitter: @d_soto_b
http://dx.doi.org/10.1016/j.tics.2016.12.009 References 1. Dehaene, S. and Changeux, J.P. (2011) Experimental and theoretical approaches to conscious processing. Neuron 70, 200–227 2. King, J.-R. et al. (2016) Brain mechanisms underlying the brief maintenance of seen and unseen sensory information. Neuron 92, 1122–1134 3. Soto, D. and Silvanto, J. (2014) Reappraising the relationship between working memory and conscious awareness. Trends Cogn. Sci. 18, 520–525 4. Rosenthal, C.R. and Soto, D. (2016) The anatomy of nonconscious recognition memory. Trends Neurosci. 39, 707–711 5. Stokes, M.G. (2015) ‘Activity-silent’ working memory in prefrontal cortex: a dynamic coding framework. Trends Cogn. Sci. 19, 394–405 6. Stein, T. et al. (2016) Can working memory be non-conscious? Neurosci. Conscious. 2016, niv011 7. Cichy, R.M. et al. (2016) Similarity-based fusion of MEG and fMRI reveals spatio-temporal dynamics in human cortex during visual object recognition. Cereb. Cortex 26, 3563–3579 8. LaRocque, J.J. et al. (2014) Multiple neural states of representation in short-term memory? It's a matter of attention. Front. Hum. Neurosci. 8, 5 9. Pan, Y. et al. (2014) Working memory biasing of visual perception without awareness. Atten. Percept. Psychophys. 76, 2051–2062 10. Baddeley, A. (1993) Working memory and conscious awareness. In Theories of Memory, pp. 11–28, Lawrence Erlbaum Associates
Trends in Cognitive Sciences, February 2017, Vol. 21, No. 2
65