Biological Cybernetics Biol. Cybernetics26, 45-52 (1977) @ by Springer-Verlag 1977 A Model of Visual Development T. Nagano InformationSciencesDivision,ElectrotechnicalLaboratory,Tokyo,Japan Abstract. There is now abundant evidence that the structure o f the mammalian visual cortex is not innately determined but can be altered by visual experiences during a certain period at an early age. A model based on the evidence is proposed that will explain some aspects of the developmental process in the visual cortex. The model self-organizingly forms "a simple cell" which responds to bars and edges of a certain orientation and retinal position. The total system o f the model corresponds to "a hyper column" which contains a complete cycle of orientational selectivity. The model has the following three original points, i) "A hyper column" is finally formed in the model, ii) Most of the cells do not have orientation selectivity in the initial state, iii) It is hypothesized that the orientation continuity of the hyper column is caused by similar orientations of successive input stimuli. The results of computer simulation show that the model has the expected performance. 1. Introduction Neurophysiological studies on the mammalian visual system have been steadily progressing since Hubel's and Wiesel's works (1962, 1963a). Many nervous network models of the visual system based on these studies have also been reported (e.g., Deutsch, 1966; Hiwatashi and Osada, 1967; Fukushima, 1969). Hubel and Wiesel (1963a) found cells responding selectively to bars or edges of a certain orientation in striate cortex of very young, visually inexperienced kittens. They concluded that the nervous networks detecting bars or edges were innately determined. This notion has been adopted in modeling of the nervous system of the brain. That is, basic features of input patterns are assumed to be first extracted by ad hoc networks with innately determined structures, and then higher-order information processings, such as concept formation by integrating the extracted features, relating concepts with each other and so on, are made by nervous networks organizable according to experiences after birth (Mart, 1969; Fukushima, 1973; Nagano, 1974). But several neurophysiological findings against this Hubel's and Wiesel's conjecture have recently been reported. They show that the visual feature extraction mechanism will also be organized according to experiences after birth (Hirsch and Spinelli, 1970; Blakemore and Cooper, 1970 ; Pettigrew and Freeman, 1973; Blakemore and Van Sluyters, 1975; Barlow, 1975). A model of visual development based on these findings are proposed in this paper. The model is simulated on a digital computer to confirm its performance. 2. Visual Development and Columnar Organization Hubel and Wiesel (1963a) showed the results that responses of neurons in the visual cortex of visually inexperienced kittens were very similar to those of adults. They concluded that feature extraction networks of the visual cortex were innately determined. But recently several findings against their conclusion are reported. Hirsch and Spinelli (1970) investigated the visual cortex of kittens that were raised from birth with one eye viewing horizontal lines and one eye viewing vertical lines. They found that elongated receptive fields of cells in the visual cortex were horizontally or vertically oriented. No oblique fields were found. Moreover, cells with horizontal receptive fields were activated only by the eye exposed to horizontal lines; cells with vertical receptive fields only by the eye exposed to vertical lines. Blakemore and Cooper (1970) reported similar results. Pettigrew and Freeman (1973) found that cells in the visual cortex of kittens raised from birth in an environment consisting entirely of small spots of a certain size responded best to the spots 46 +'-kkttlt+ --.-',, ',, f / / . - ' - --.--...',,\ttt...'----'----\ k t t t --'--Fig. 1. Orientation selectivityof E-cellsin a hyper column. Each bar indicates the optimal orientation of a given E-cell . . \,,,k dimensions. The model proposed here deals only with the simple cell as is so in all the other models of visual development. Only the orientation selectivity of cells is considered and ocular dominance of cells is not taken into account to simplify the model. 3. Review of Visual Developmental Models Models of visual development proposed till now are summarized in order to clarify the differences between them and the model proposed here. / #-.- '" Von der Malsburg's Model (1973) "--" / t " - " - " t t \ ' , . 9 " " t/ 9 # "~'~,-~,,~ Fig. 2. View onto the cortex after 100 steps of learning. Each bar indicates the optimal orientationo f a givenoutput cell. Dots without a bar are cellswhichnever reactedto the standard set of stimuli (yon der Malsburg, 1973) with the same size. It is certain from these findings that visual experiences during a certain period at an early age change the visual nervous system. A finding supporting this notion has also been reported in the study of binocular vision (Wiesel and Hubel, 1963). It is well known that the neocortex consists of an array of columns each of which corresponds to some functional unit (Lorente de No, 1938). In the visual cortex each column consists of cells responding selectively to lines and edges of a certain orientation, and neighboring columns have similar orientation selectivity (Hubel and Wiesel, 1963b, 1968). Recent Hubel's and Wiesel's works (1974a, 1974b) show that the visual cortex has a higher-order functional unit called hypercolumn. A hypercolumn is, the area of cortex about 1 or 2 mm square, within which a complete cycle of ocular dominance is contained when moving across the cortex in one direction, and a complete cycle of orientational selectivity when moving in the orthogonal direction (see Fig. 1). Hyper columns are arranged like a mosaic on the cortex. As is generally known, the visual cortex has roughly three types of cells : the simple cell responding to lines and edges of a certain orientation and retinal position, the complex cell responding to lines and edges of a certain orientation on a wide area of the retina, and the hyper complex cell responding selectively to bars or edges with particular orientations and various critical His model is composed of the uniform two dimensional array of pairs of an excitatory cell (E-cell) and an inhibitory cell (I-cell). Each afferent input fiber connects only to E-cells via excitatory modifiable synapses. All the connections between cells in the cortex are made with fixed synapses. Each E-cell connects to other Ecells in its surrounding region of a certain size. Each Icell connects to E-cells in its surrounding region of a larger size than that of the E-cell's region. No connections are assumed between/-cells. Initial values of synaptic efficacies between input fibers and E-cells are random. Each synapse changes its efficacy only when the post synaptic E-cell is active. It increases its efficacy when its input fiber is active and decreases its efficacy otherwise. Synaptic efficacies of each E-cell are normalized to keep the summation of synaptic efficacies constant whenever efficacies are modified. Each E-cell is organized to respond to lines of paticular orientation by the iterative presentation of lines with various orientations. Figure 2 is a result of computer simulation. It shows that this model well simulates the formation of columnar organization since the E-cells with the same orientation selectivity form a cluster. Nass's and Cooper's Model (1975) Their model is similar to yon der Malsburg's one except that their model lacks connections between E-cells. A model with/-cells is compared with a model without Icells. It is shown that/-cells are necessary for E-cells to be organized to have orientation selectivity. The constancy of the summation of synaptic efficacies is not assumed because of its improbability in the actual nervous systems. Instead of the restriction, the decay of each efficacy proportional to its value is introduced throughout the learning period to produce similar learning effect without the restriction. It is shown that the concentric receptive field organization of lateral geniculate cells makes the output of an E-cell minimum when input bars are a little rotated fl'om its optimum 47 orientation. They propose that this effect may provide a cue signal for neighboring columns to be organized to have similar orientation. Hall's and Yau' s M o d i f i c a t i o n A l g o r i t h m (1976) o|174 %o| 9 The way of modifying efficacies (modification algorithm) to produce orientation selectivity is studied mathematically. In order to organize E-cells to have orientation selectivity without lateral inhibition by Icells, two types of modification algorithms are necessary: one is Hebb's way that increases synaptic efficacy only when simultaneous excitation of pre- and post synaptic cells occurs and otherwise decreases it gradually; the other is the way similar to that used in vonder Malsburg's model except the restriction that the summation of efficacies is constant. The former way is necessary at the early stage of learning and the latter at the later stage of learning. Grossberg (1976) also has proposed a model of visual development. A lateral inhibition mechanism other than Nass's and Cooper's one is proposed in his model. It enables his model to detect only relative intensities of inputs. The restriction that the summation of efficacies is constant can be eliminated by using relative intensities of inputs as input signals. 4. Description of the Model Three original points of the model are first described. i) Almost all the cells have no orientation selectivity in the initial state. ii) A hyper column shown in Figure 1 is formed. iii) The formation process of a hyper column depends on input pattern sequence. The review of neurophysiological findings in 2 shows that cells of young visually inexperienced kittens will have weak orientation selectivity and that their "tuning" characteristics will be improved by visual experiences after birth. It may, therefore, be accepted that E-cells of a model have weak orientation selectivity in its initial state as in yon der Malsburg's model. But a model with such an initial structure is not interesting because it is easy to devise such a model that is composed of E-cells with innately determined orientation selectivity and that only tuning characteristics of E-cells are improved by visual experiences. Almost all the E-cells of the model proposed here are, therefore, assumed to have no orientation selectivity in its initial state. Only a very few E-cells have weak orientation selectivity in order to trigger the model because the model could not begin to form orientation selective cells if no cells had orientation selectivity. 9 9 %%...% Fig.3. Arrayof cellsin the presentmodel. 9and @are an E-celland an /-cell, respectively All the previous models reviewed in 3 simulate only the formation of the columnar organization and the improvement of orientation selectivity. None of them do not simulate the hyper columnar organization in Figure 1. As was previously described, Nass and Cooper considered that a large falling-off of E-cell's output caused by the concentric receptive field organization of lateral geniculate cells might become a cue signal for neighboring columns to be organized to have similar orientation selectivity. According to their thought, the orientation difference between two neighboring columns becomes about 20 ~. But the difference in the actual cortex is far smaller, [e.g., 7.2~ ~ in the monkey striate cortex; Hubel and Wiesel (1974a)]. It does not, therefore, seem that the actual cortex uses such a cue signal. The continuity of columnar organization is assumed to be caused by continuous changes of input orientation in the model proposed here. The adequency of this hypothesis is as follows. Newborn animals seem to look their environments around uncouciously without gazing some paticular things intentionally. Their heads and bodies seem to move slowly and continuously. It is, therefore, quite natural to think that input bars and edges are presented to the retina with continuous changes of their orientations. It also supports this hypothesis that cells with each orientation selectivity are equally dense in spite of the predominant majority of vertical and horizontal lines to oblique ones in the actual environments. The structure of the model is described in the following. The model is composed of M x N two dimensional array of pairs of an E-cell and an/-cell as is shown in Figure 3. Each E-cell is the output cell of the cortex and/-cells are interneurons. Input signals from the lateral geniculate cells are denoted by X =(Xl, x2,..., xL), where o < x t < Xmax(l = 1, 2 , . . . , L). (1) 48 Xl X 2 X3 9 9 , Table 1. Notations of synaptic efficacies XL Connection Synaptic efficacy xl-~ E-cell E-cell~ E-cell E-cell--,I-cell /-cell-~ E-cell wz.ij(1 = 1, 2 , . . . L ; i = 1, 2 , . . . M ;j = 1, 2 , . . . N ) amj, u(m + i; m, i = 1, 2, ...M ;j = 1, 2,...N) b,~j. u(m, i = 1, 2, . . . M ;j = 1, 2 , . . . N ) c,,,, u(m, i = 1, 2, .. . M ; n ~ j ; n,j = 1, 2, . ..N) Fig. 4. Connection between input fibers and an E-cell. ~ indicates an excitatory modifiable synapse. Every input fiber connects to each E-cell via an excitatory modifiable synapse O O & 0 O (30 0 0 Fig. 5. Connection between E-cells. --<3 indicates a excitatory fixed synapse. Each E-cell connects to the other E-cells in the same column C) 0 Fig. 6. Connections between E-cells and an/-cell. ~ O indicates an inhibitory fixed synapse. Each E-cell receives inputs only from all the E-cells in the same column via excitatory fixed synapses. Each/-cell connects to the E-cells in the other columns via inhibitory fixed synapses Every input fiber connects to each E-cell via an excitatory modifiable synapse (Fig. 4). No input fibers connect to/-cells directly. Connections between E-cells are shown in Figure 5. Only E-cells in the same vertical column of the array are connected with each other via excitatory fixed synapses. Figure 6 shows connections between E-cells and an/-cell. All the E-cells in the same column connect to each /-cell in that column via excitatory fixed synapses. Each /-cell connects to Ecells in all the other columns via inhibitory fixed synapses, though Figure 6 shows only connections between neighboring columns for simplicity. Synaptic efficacies of connections between neighboring columns have to be maximum. No connections are assumed between/-cells. Input output relations of the two types of cells are described. The summation of all the weighted inputs of a cell is called the "potential" of the cell. Potentials of Ecells and /-cells are denoted by yij(t) and zu(t ) (i = 1, 2,..., M ; j = 1, 2,..., N), respectively, where t is time. Output of E-cells and/-cells are denoted by y*(t) and z*(t), respectively. The relation between a potential and an output is given by Figure 7. Notations Ymax,*Zmax,*0y, 0z, Ys, and zs are easily understood from this figure. All these parameters are assumed to be independent of i, j. Notations of synaptic efficacies are defined in Table 1. Potentials and outputs of the E-cells and the/-cells are given by the following equations. L M y,j(O = Y~ w~,,j(Oxz(O+ ~ amj,iy*~(t-- t) l=1 m=l m*i M - Z N Z cm.,uz*.(t- lt, (2) ra=l n = l n~j y orz (3) y*(t) = Oy(yu(t)) , (4) z~(t) =o~(zij(t)) , (5) m=l Yrnax or Z~ rnox 0 M Zij(t)= E bmj, iJYmj(t), @ o r 8z y~ orz~ y orz Fig. 7. Input-output relations of an E-cell and an/-cell where Oy and O~ are the analog threshold functions defined by Figure 7. The modification of synaptic efficacy is assumed to occur only when the potential of an E-cell exceeds the modification threshold d, where d is greater than the output threshold 0r The modification equation in this 49 model is given by wl, ij(t) = wt, ij(t - 1) + k 1xt( t - 1)y~:( t - 1) XI(I.- 1))y/*j(t- 1), --]s (6) where y~j(t) is defined by y[j(t) = y i i ( t ) - d , = 0, if y,j(t)>d otherwise, (7) k~ >0, k 2 > 0 and 0<Wl,~j_--<Wma~. This modification algorithm is similar to the second one in Hall's and Yau's work except the existence of the modification threshold which is greater than the output threshold 0y. 5. Initial Conditions and Principles of Behavior Three assumptions are made for simplification. i) x t = 1 or 0, (1= 1, 2, ..., L). ii) IXl = K = const, where IXl means the number of components whose values are 1. iii) Each/-cell connects only to E-cells in the two neighboring columns. In order to satisfy the condition that ahnost all the cells have no orientation selectivity in the initial state, synaptic efficacies amj,~i, b~j,~j, and cm;+t,ij are made independent of m, i, andj. These constant efficacies are expressed by a, b, and c, respectively. Almost all the modifiable synaptic efficacies have the same initial value Wo(0< w o < W~a~). The rest (which are very few) have a value w o + A w o ( A w o >0). As the formation of orientation selectivity is caused by the differences of the potentials between E-cells, the model could not start if all the E-celIs had an equal potential. A very few synaptic efficacies are, therefore, given a greater initial value than others to produce a small potential difference which triggers the model. The initial values of synaptic efficacies make almost all the E-cells have equal potential to every input pattern (a bar or an edge). A very few E-cells with the synaptic efficacy w o + A w o have a different potential when the synapse receives an active input signal. The following roughly explains how a hyper column is formed from these initial conditions. An E-cell having a synapse with efficacy w o + A w o is denoted by E 0. An input pattern supplying an active signal to the synapse is denoted by X o. The potential of E o is higher than other E-cells by A w o when X 0 is presented. This potential difference causes the change of synaptic efficacies of E o according to (7) so as to produce the selectivity to X o. The potential difference is transmitted to other E-cells in the same column of the array via colateral fibers in Figure 5, resulting in the formation of the same orientation selectivity for the E-cells. As is understood from (7), the E-cells organized to have the selectivity to X o have lower potential than the initial potential for any other input pattern with different orientation. This falling-off of potentials is transmitted to the E-cells in the neighboring columns v i a / - c e l l s in Figure 6, resulting in the increase of the potentials of the E-cells in the neighboring columns. The synaptic efficacies of these E-cells are, therefore, modified so that the E-cells may have the selectivity to the input pattern X 1 following X o. Consequently a hyper column is formed if the orientation of input patterns change gradually and continuously. Conditions for a hyper column to be formed is derived in the following. It is necessary that the potential change of each cell causes the change of its output in the initial state, The initial potentials yi~(0), zij(0), therefore, have to satisfy 0y < y,.j(0)< Ys, 0~ < z~j(0)< z s . (8) In order for the model to be triggered only by the input to the synapse with the efficacy w o + A w o , the potentials of all the E-cells are lower than the modification threshold d when such synapses are not activated. In such a case, y~(0) and z~j(0) are independent of ij and are denoted Yo and zo, respectively. From (2) and (3), Yo and z o are given by Yo = K w o + ( M - 1)avy(yo - Oy) - 2mcTz(Zo - Oz), z o = M b T , ( y o - 0,), (9) (10) where 7y and 7z are the gradients of the linear parts of the relations in Figure 7. Hence we have Yo = K w o - ( M - l)aTy0y + 2M2bcyyTzO~, + 2 M c ~ O ~ < d. 1 - (M - 1)avy + 2M2bcTy7z (11) The condition for the synaptic efficacies of E o to be modified by the presentation o f X 0 in the initial state is given by yo + A w o > d . (12) During the presentation of each input pattern, h times of modifications are assumed to be made. Let 5w+ and cSw_ be, respectively, the increase and the decrease of synaptic efficacies of E o caused by X o. Ignoring the effect that the potential change of E o is fed back to E o via other E-cells by assuming a, b, c ~ 1, we have c%v+ = kick(1 + k l K ) h- 1, (13) ~w_ = -- k2~(1 + kaK) h- 1, (14) where c~=yo + A wo - d. (15) 50 TINES OF ITERATION 0000000000Q 00000000000000 0000 0 0 0 0 0 O0 00000000000000 00000 000000 1 2 3 4 = 5 0 00 TIMES OF ITERATION 12300000000032 12300000000032 12300000000032 12300000000032 12300000000032 TIMES OF ITERATION = 2 0 0 00 0 0 0 0 0 00 0 00 0000 00000000 O0 00000 00 00 O 0 0 0 0 00000000000000 00 00 0 0 0 00 00 0 00 TINES OF ITERATION 12345600065432 12345600065432 12345600065432 12345600065432 12345600065432 - 6 TIMES OF ITERATION = 3 10000000000000 000000000 O0 0 0 0 00000 000000 000 00000000000000 0000 0000 000000 TIMES OF ITERATION 12345670765432 12345670765432 12345670765432 12345670765432 12345670765432 = 7 TIMES OF ITERATION = 8 TIMES OF ITERATION : 1 000 0 00 : 4 12000000000002 5 6 7 8 Fig. 8. Straight line segments used as input patterns in the computer simulation TINES OF ITERATION 00000000000000 00000000000000 00000000000000 00000000000000 00000000000000 = 1 TIMES OF ITERATION I0000000000000 00000000000000 00000000000000 00000000000000 00000000000000 : 2 TIMES OF ITERATION 12300000000032 12300000000032 12300000000032 12300000000032 12300000000032 = 3 TIMES OF ITERATION 12345600065432 12345600065432 12345600065432 12345600065432 12345600065432 = 4 TIMES OF ITERATION = 5 12000000000002 12000000000002 12000000000002 70 765432 781765432 781765432 2345678[765432 Fig. 12. Formation process of a hyper column in the case of partially irregular input sequence. The numerals and the square in this figure have the same meanings as those in Figure 9 The condition for the potentials o f the E-cells in the same c o l u m n as E o to exceed the modification threshold on the presentation of X o is given by Yo + a?y(Awo+ K6w +)> d. (16) The number o f input fibers which are activated by both X o and X 1 is denoted by f. The condition for the potentials o f the E-cells in the neighboring columns to exceed the modification threshold on the presentation of X 1 is given by yo-C?~{bTy(f6w+ + ( K - f ) 6 w +Awo) } > d . 7 ~ 7 6 5 4 3 2 7~765432 234567~765432 Fig. 9. Formation process of a hyper column in the case of systematic input sequence. The part enclosed by a square is a hyper column. The numerals in this figure indicate each orientation of input patterns (see Fig. 8). The numeral 0 means that the potential of a given E-cell is not high enough for any one of the input patterns (17) We can assume f ~ K since X o and X 1 are straight line segments with different orientation. The second term of the left side of (17) can, therefore, b e c o m e positive so that (17) m a y be satisfied, i f k 1 and k 2 are not so m u c h different. 6. Computer Simulation POTENTIAL DISTRIBUTION 1 . 0 0 1 . 0 0 1 . 0 0 1 . 0 0 0 . 6 0 3 . 4 0 2 . 7 0 9,91 1 . 0 0 1 . 0 0 1 . 0 0 l . O0 0.60 3.40 2.70 9.91 1.00 1.00 1.00 1.00 0.60 3.40 2.10 9.91 1 . 0 0 1 . 0 0 1 . 0 0 1 . 0 0 0 . 6 0 3 , 4 0 2.70 9.91 1.001,001.001,000.603.402.709.91 2 . 7 0 3 . 4 0 0 . 6 0 l . O01.001.00 2.703.400.601.001.001.00 2.70 3.40 0.60 1.00 1,00 1.00 2.70 3 . 4 0 0 . 6 0 1 . 0 0 1 . 0 0 1 . 0 0 2.703.400.601.001.001.00 Fig. 10. Potential distribution for the pattern No. 8 after five times of interation WEIGHT OISTRIBUTION 0. 0. O. O. 1.000 O. O. O. O. O. O. : O. 1.000 O. O. O. O. O. Q. O. 1.000 O. O. 0. O. Oo O. O. 1.000 O. O. O. O. O. O. O. 1.000 0. O. O. O. O. O. O. 1.000 O. O. O. O. O. O. O. 1.000 O. O. O. O. 0. O. O. O. O. O. O. O. O. O. 1.000 O. 0. O. O. O. O. O. O. I,000 O. O, O. O. Fig. 11. Final synaptic efficacies of an E-cell responding selectively to the pattern No. l Eight straight line segments with different orientations shown in Figure 8 were used as input patterns. Black meshes and white meshes of each pattern are assumed to express active inputs (1) and silent ones (0), respectively. The array of cells are 5 x 14, that is, M = 5 and N = 14. The first c o l u m n and the 14-th column are connected with the same way as other neighboring columns in order to m a k e the structure of a small-scale m o d e l uniform. Specific values o f other parameters are amj, ij = 0.1(m @ i;j= 1, 2, ..., 14; m, i = 1, 2, ..., 5) b,,j,u = 0.1(j = 1, 2 , ..., 14; re, i = 1, 2, ...,5) %-, u = 0.08(n = j + 1 ;j = 1, 2 , . . . , 14 ; m, i = 1, 2 . . . . . 5) wl, q(0) = 0.5(/= 1, 2 . . . . . 81 ;j = 1, 2 . . . . , 1 4 ; i = 1, 2,..., 5) 51 with the exception that w37' it(0) = 1.0, Oy = 0 z = 0 . 0 ; d=4.5 ; Y~ = Zs = ~ kl =k2=0.1 ; ; 7y = ~ = 1 . 0 ; Wmax=Xmax=l.0. Input patterns were iteratively presented in order according to the numbers of the patterns in Figure 8 to satisfy the hypothesis about the way of input pattern presentation. Figure 9 shows the numbers of the patterns which are selected by each E-cell. A hyper column (a part enclosed by a square in this figure) is formed after five times of iteration. Figure 10 shows the potential of each E-cell caused by the pattern No. 8 after learning. Figure 11 shows the final synaptic efficacies of an E-cell responding selectively to the pattern No. 1. Both figures show that the model has the expected performance. Each pattern was presented completely in order in the above simulation experiments so that the orientation of input patterns may change continuously from 0 ~ to 180 ~ But large and distrete changes of orientations seem to occur sometimes on the actual retina of newborn animals. The performance of the model to such an input sequence was also investigated. The input sequence in this case was composed of the iteration of a cycle 123456, 345678, 567812, 781234. Figure 12 shows the result that the formation of a hyper column was successful even when input sequence contained sudden changes of orientations. 7. Discussion The following three facts are known about the neocortex: the number of the principal neuron (E-cell) is roughly equal to that of the interneuron (/-cell); there exsist inhibitory interneurons; axon colaterals of a principal neuron are connected to other principal neurons. Though the way ofaxon branching is not clear about inhibitory interneurons of the neocortex, inhibitory interneurons in the other parts of the brain (e.g., the cerebellar cortex and the hippocampus) extend their axon branches only to a particular direction. The structure o f the model does not conflict with these facts. One of the original points of this research is the hypothesis that the continuity of the columnar organization is caused by the continuous change of the orientation of input patterns. The continuity of the columnar organization does not seem to be indispensable for further in formation processing after the extraction of line segments. On the other hand, it is very difficult to devise a model which realize the continuous columnar organization regardless of input pattern sequences. Such a model seems to have a structure with great complexity caused only by the realization of the continuity. Such an unbalance of the little necessity and the great complexity does not seem to exist in living organisms since they are considered to be reasonably organized by the natural selection. This is the reason why the continuous columnar organization is assumed to be formed simply by the continuity of the orientation. As this hypothesis can be examined by neurophysiological experiments, this model can be said to offer a research theme to experimental neurophysiologists. Acknowledgements. The author thanks to Mr. T. Hamada and Dr. B. Furubayashi of Electrotechnical Laboratory for their helpful discussions. References Barlow, H.B.: Visual experience and cortical development. Nature 258, 199 204 (1975) Blakemore, C., Cooper, G. F. : Development of the brain depends on the visual environment. Nature (Lond.) 228, 477 478 (1970) Blakemore, C., Van Sluyters, R. C. : Innate and environmental factors in the development of the kitten's visual cortex. J. Physiol. 248, 663--716 (1975) Deutsch, S.: Conjectures on the mammalian neuron networks for visual pattern recognition. IEEE Trans., SSC-2, 81--85 (1966) Fukushima, K.: Visual feature extraction by a multinetwork of analog threshold elements. IEEE Trans., SSC-5, 322 333 (1969) Fukushima, K.: A model of associative memory in the brain. Kybernetik 12, 58--63 (1973) Grossberg, S. : On the development of feature detectors in the visual cortex with applications to learning and reaction-diffusion systems. Biol. Cybernetics 21, 145--159 (1976) Hall, R., Yau, S. S. : The distribution of orientation of optimal stimuli for cells of striate cortex. Biol. Cybernetics 21, 113 120 (1976) Hirsch, H.V.B., Spineili, D.N.: Visual experience modifies distribution of horizontally and vertically oriented receptive fields in cats. Science 168, 869--871 (1970) Hiwatashi, K., Osada, S. : A simulation of receptive fields in the vertebrate visual system by electronic circuits. Jap. J. Med. Electron. Biol. Eng. 5, 376 383 (1967) Hubel, D.H., Wiesel, T.N.: Receptive fields, binocular interaction, and functional architecture in the cat's visual cortex. J. Physiol. (Lond.) 160, 106--154 (1962) H ubel, D. H., Wiesel, T. N. : Receptive fields o f cells in striate cortex o f very young, visually inexperienced kittens. J. Neurophysiol. 26, 994--1002 (1963a) Hubel, D.H., WieseI, T.N.: Shape and arrangement of columns in cat's striate cortex. J. Physiol. 165, 559--568 (1963b) Hubel, D.H., Wiesel,T.N. : Receptive fields and functional architecture of monkey striate cortex. J. Physiol. (Lond.) 195, 215--243 (1968) Hubel, D.H., Wiesel, T.N.: Sequence regularity and geometry of orientation columns in the monkey striate cortex. J. Comp. Neurol. 158, 267--294 (1974a) Hubel, D.H., Wiesel, T.N.: Uniformity of monkey striate cortex: a parallel relationship between field size, scatter and magnification factor. J. Comp. Neurol. 158, 295--306 (1974b) Lorente de No, R.: The cerebral cortex: Architecture, intracortical connections, and motor projections in physiology and nervous system. New York : Oxford University Press 1938 52 Marr, D. : A theory for cerebral neocortex. Proc. Roy. Soc. Lond. B 176, 161-234 (1970) Nagano, T. : Some considerations on learning algorithms. Trans. Inst. Electron. Comm. Eng. Jap. 57-C, 661--667 (1974) Nass,M.M., Cooper, L.N. : A theory for the development of feature detecting cells in visual cortex. Biol. Cybernetics 19, 1--18 (1975) Pettigrew,J.D., Freeman, R.D.: Visual experience without lines: effect on developing cortical neurons. Science 182, 599--601 (1973) Shepherd,G. M. : The synaptic organization of the brain. New York: Oxford University Press, 1974 Von der Malsburg, C. : Self-organization of orientation sensitive cells in the striate cortex. Kybernetik 14, 85--100 (1973) Wiesel, T.N., Hubel, D.H.: Single-cell response in striate cortex of kittens deprived of vision on one eye. J. Neurophysiol. 26, 1004~ 1017 (1963) Received: November 25, 1976 Information Sciences Division Electrotechnical Laboratory 5-4-1, Mukodai-machi, Tanashi-shi, Tokyo, 188, Japan