Lateralization in a bihemispheric neural model of letter identification

Lateralization in a bihemispheric neural model of letter identification

Neurocomputing 26}27 (1999) 875}880 Lateralization in a bihemispheric neural model of letter identi"cation Natalia Shevtsova , James Reggia * Inst...

186KB Sizes 0 Downloads 40 Views

Neurocomputing 26}27 (1999) 875}880

Lateralization in a bihemispheric neural model of letter identi"cation Natalia Shevtsova , James Reggia * Institute for Advanced Computer Studies, University of Maryland, College Park, MD 20742, USA Institute for Advanced Computer Studies, Departments of Computer Science and Neurology, University of Maryland, College Park, MD 20742, USA

Abstract The causes of cerebral lateralization of cognitive and other functions are currently not well understood. A bihemispheric neural network model for a simple visual identi"cation task was developed to study this issue, with two hemispheric regions interacting via a simulated corpus callosum. Lateralization occurred when underlying hemispheric asymmetries were present, including asymmetric size, excitability, or synaptic plasticity, supporting the belief that biological lateralization may be caused by multiple factors independently. Lateralization was more intense with strong inhibitory callosal connections, supporting the hypothesis that the corpus callosum plays a functionally inhibitory role.  1999 Elsevier Science B.V. All rights reserved. Keywords: Neural network; Lateralization; Corpus callosum

1. Introduction While hemispheric specializations for di!erent behavioral and perceptual tasks are well documented, the underlying causes of lateralization of these functions are not completely understood and remain a subject of intense study. There is experimental evidence that corresponding cortical regions can di!er in their relative size [4], dendritic branching [10], neurotransmitter levels [13], and excitability to external stimuli [7], but which of these or other underlying asymmetries leads to lateralization is currently unclear. The corpus callosum (neural connections between the hemispheres)

* Corresponding author. Tel.: #1-301-405-2686; fax: 301-405-6707; e-mail: [email protected].  Present address: A.B. Kogan Research Institute for Neurocybernetics, Rostov State University, 194/1 Stachka Ave., Rostov-on-Don 344090, Russia. 0925-2312/99/$ } see front matter  1999 Elsevier Science B.V. All rights reserved. PII: S 0 9 2 5 - 2 3 1 2 ( 9 8 ) 0 0 1 3 2 - 5

876

N. Shevtsova, J. Reggia / Neurocomputing 26}27 (1999) 875}880

may also be a potential factor in function lateralization. However, at present there are disagreements in the literature about whether interhemispheric interactions are predominantly excitatory or inhibitory [1,2,6,8].

2. Model description We created and studied a bihemispheric neural network model for a simple visual information processing task (letter identi"cation). Fig. 1 shows the model structure schematically. The input layer is separated into left and right visual hemi"elds (LVF and RVF). Each letter stimulus may be presented in the LVF, RVF, or center of the visual "eld. Each hemi"eld projects onto the contralateral primary cortical layer. The orientation-sensitive neurons in the primary cortical layers extract orientation of local edges in their receptive "elds. Each input stimulus is encoded in distributed activity over the primary layers. Each primary layer projects onto the corresponding associative cortical layer. The left and right associative layers interact via simulated homotopic callosal connections. Lateral intracortical connections in the associative layers, callosal connections, and unsupervised learning of primary-to-associative

Fig. 1. Architecture of the bihemispheric neural model. VM } vertical meridian, LVF } left visual hemi"eld, RVF } right visual hemi"eld.

N. Shevtsova, J. Reggia / Neurocomputing 26}27 (1999) 875}880

877

connection strengths generate an encoding of the input stimuli over the associative cortical layers. Both associative cortical layers project to an output layer, which is a one-dimensional layer of linear neurons. Each output element represents a letter in a certain position in the visual "eld. Supervised learning is used to modify associative-to-output connection strengths during training. Model performance is measured as root mean square error E of output element responses in terms of correctly classifying each input stimulus.

3. Experimental methods Using this basic model, we performed a series of computer experiments to examine function lateralization with di!erent assumptions about system parameters. Each simulation involved training the model to recognize a set of letters presented as input stimuli. System performance was studied as a function of assumed callosal connection strength. A version of the model with symmetric left and right hemispheres served as a control. Di!erent types of asymmetry were studied one at a time: size of associative cortical layers, excitability in the associative layers, and rates of unsupervised and supervised learning. The pretraining value of E for the model ranged from 5 to 8. In each simulation, training continued until either E was reduced to 0.05 or 1000 presentations of the training stimuli occurred. In most cases, after training the model was able to identify the input stimuli correctly (100% correct performance). Lateralization was measured as the di!erence in contribution of the two hemispheres to performance. Speci"cally, after training was completed, the root mean square error was measured under three conditions: with both associative layers connected to output layer (E) and with each of the left and the right associative cortical layers alone connected to output elements (E* and E0, respectively, Fig. 2). Lateralization was measured as E*!E0. Negative values correspond to left lateralization, positive values to right.

Fig. 2. Individual left (E*) and right (E0) hemispheric errors are measured with only that hemisphere connected to outputs (double horizontal lines indicate blocked outputs).

878

N. Shevtsova, J. Reggia / Neurocomputing 26}27 (1999) 875}880

4. Simulation results In the symmetric model, which served as a control, all parameters were identical in both hemispheres except for the initial random primary-to-associative and associative-to-output connection weights. Lateralization following training is very close to zero for all values of callosal connection strengths, being both slightly positive or

Fig. 3. Post-training lateralization (solid black line) and initial mean activation in the left (dashed gray line) and right (solid gray line) associative cortical layers vs. callosal strength. (a) Symmetric case. (b) More excitable left than right associative cortical layer neurons. (c) Larger left than right associative cortical layers. (d) Same as (c) except with adjusted callosal strength to eliminate di!erence in initial activation levels. (e) Left larger than right unsupervised learning rate. (f ) Left larger than right supervised learning rate.

N. Shevtsova, J. Reggia / Neurocomputing 26}27 (1999) 875}880

879

negative in di!erent simulations (Fig. 3a). The direction and value of lateralization are determined by small di!erences in the initial random weights. Asymmetric cortical excitability is a potential causative factor for lateralization [7]. In the model, asymmetric cortical excitability leads to asymmetric mean activation levels, and a higher initial mean activation on the left is accompanied by lateralization to the left (Fig. 3b). This is more pronounced for inhibitory callosal connections. Another hypothesized cause of lateralization is the di!erence in size of cortical regions in the opposite hemispheres [3,9]. The results for a four-fold size asymmetry favoring the left are shown in Fig. 3c. For inhibitory callosal connections, a di!erence in initial mean activation in the hemispheric regions due to size asymmetry may itself lead to lateralization. To eliminate this factor, we adjusted the callosal connection strength from the left associative layer to the right one, so mean activations were roughly equal (Fig. 3d). Another potential cause of lateralization is a di!erence in neurotransmitter concentrations in the left and right cerebral hemispheres [12]. Such a di!erence might be related to a di!erence in synaptic plasticity in the two hemispheres. In the model, asymmetric synaptic plasticity was simulated by having di!erent learning rates for primary-to-associative (unsupervised learning) and associative-to-output (supervised learning) connection weight changes (Fig. 3e and f, respectively). Representative examples of simulation results with the model for two values of callosal strength (inhibitory callosal connections, c"!4, and excitatory callosal connections, c"#1) are presented in Table 1. Table 1 Asymmetry

Callosal strength

RMSE

Lateralization

Full

Left

Right

Mean activation Left

Right

Symmetric

!4 #1

0.049 0.049

0.302 0.352

0.289 0.374

0.013 !0.022

0.145 0.276

0.162 0.278

Excitability k"1/0.3

!4 #1

0.050 0.082

0.173 0.481

0.376 0.588

!0.203 !0.107

0.208 0.241

0.034 0.162

size 20;20/10;10

!4 #1

0.050 0.049

0.186 1.109

0.362 0.984

!0.175 0.124

0.208 0.330

0.095 0.502

Size 20;20/10;10

!4 #1

0.049 0.050

0.290 0.918

0.359 0.849

!0.069 0.069

0.158 0.270

0.169 0.267

Learning rates 0.01/0.001 (unsupervised)

!4 #1

0.050 0.049

0.192 0.458

0.356 0.384

!0.164 0.074

0.205 0.255

0.046 0.195

Learning rates 0.001/0.0001 (supervised)

!4 #1

0.050 0.050

0.382 1.414

0.438 1.382

!0.056 0.032

0.145 0.239

0.162 0.239

Activation balancing done; see [11].

880

N. Shevtsova, J. Reggia / Neurocomputing 26}27 (1999) 875}880

5. Discussion The results indicate that multiple asymmetries may cause lateralization. Lateralization occurred toward the side having larger size, higher excitability, or higher learning rate parameters. The fact that multiple underlying asymmetries can cause lateralization in the model is consistent with arguments that a single underlying factor probably does not cause human behavioral lateralization (e.g., [5]). Lateralization increased with increasing di!erence in asymmetric parameters. It appeared more intensively with strong inhibitory callosal connections, supporting the hypothesis that the corpus callosum plays a functionally inhibitory role.

References [1] N.D. Cook, The Brain Code: Mechanisms of Information Transfer and the Role of the Corpus Callosum, Methuen, London, 1986. [2] A. Ferbert, A. Priori, J.C. Rothwell, Interhemispheric inhibition of the human motor cortex, J. Physiol. 453 (1992) 525}546. [3] A.M. Galaburda, F. Sanides, N. Geschwind, Human brain: cytoarchitectonic left}right asymmetries in the temporal speech region, Arch. Neurol. 35 (1978) 812}817. [4] N. Geschwind, W. Levitsky, Human brain: left}right asymmetries in temporal speech region, Science 161 (1968) 186}187. [5] J.B. Hellige, Hemispheric asymmetry: What's Right and What's Left, Harvard University Press, Cambridge, 1993. [6] M. Kinsbourne (Ed.), Asymmetrical Function of the Brain, Cambridge University Press, Cambridge, 1978. [7] R. Macdonell, B.E. Shapiro, K.H. Chiappa, Hemispheric threshold di!erences for motor evoked potentials produced by magnetic coil stimulation, Neurology 41 (1991) 1441}1444. [8] B. Meyer, S. RoK richt, H. GraK "n von Einsiedel, F. Krudgel, A. Weindl, Inhibitory and excitatory interhemispheric transfers between motor cortical areas in normal humans and patients with abnormalities of corpus callosum, Brain 118 (1995) 429}440. [9] R.J. Nudo, W.M. Jenkins, M.M. Merzenich, T. Prejean, R. Grenda, Neurophysiological correlates of hand preference in primary motor cortex of adult monkey, J. Neurosci. 12 (1992) 2918}2947. [10] A.M. Scheibel, I. Fried, L. Paul, A. Forsythe, U. Tomiyasu, A. Wechsler, A. Kao, J. Slotnick, Di!erentiality characteristics of the human speech cortex: A quantitative Goldgi study, in: D. Benson, E. Zaidel (Eds.), The Dual Brain, Guilford Press, New York, 1985, pp. 65}74. [11] N. Shevtsova, J. Reggia, A neural network model of lateralization during letter identi"cation, J. Cogn. Neurosci. in press. [12] D.M. Tucker, Hemisphere specialization: A mechanism for unifying anterior and posterior brain regions, in: D. Ottoson (Ed.), Duality and Unity of Brain, Macmillan Press, London, 1987, pp. 180}193. [13] D.M. Tucker, P.A. Williamson, Asymmetric neural control systems in human self-regulation, Psych. Rev. 91 (1984) 185}215.