Lecture Notes in Computer Science
A Ring Model for the Development of Simple Cells in the
Download 12.42 Mb. Pdf ko'rish
|
- Bu sahifa navigatsiya:
- Keywords
- 2 Methods 2.1 Structure of the Model
- Fig. 1.
- 2.2 Correlation of Spontaneous Activities in LGN
- 2.3 Initial Synaptic Weights
- 2.4 Modification of the Synaptic Weights
- 3 Results 3.1 Development of Simple Receptive Fields
A Ring Model for the Development of Simple Cells in the Visual Cortex Takashi Hamada 1 and Kazuhiro Okada 2
1 National Institute of Advanced Industrial Science and Technology (AIST), Kansai center 1-8-31, Midorigaoka, Ikeda, Osaka 586-8577, Japan hamada-takashi@aist.go.jp 2 Murata Machinery, Ltd., Communication Equipment Division 136, Takeda-Mukaishiro-cho, Fushimi, Kyoto 612-8686, Japan kazuhiro.okada@nan.muratec.co.jp Abstract. A model was proposed for the development of simple receptive fields in the cat visual cortex, based on the empirical evidence that the development is due to dark or spontaneous activities in the lateral geniculate nucleus (LGN). We assumed that several cortical cells are arranged in a ring, with mutual excitatory and inhibitory connections of fixed weights; The cells also receive excitatory synapses from LGN cells, whose synaptic weights are initially set to be random and then modified according to the Hebb rule as well as to positive correlations among nearby LGN cells. Computer simulation showed that the cortical cells finally acquire two-dimensional simple receptive fields with their phases gradu- ally varied along the ring. Keywords: Simple cell, development, spontaneous activity, correlation, Hebb rule, model. 1 Introduction Cells in the layer 4 of the cat visual cortex have so-called simple receptive fields in which several elongated On and Off sub-areas are alternately arranged [1]. If a slit of light is positioned on the On sub-area with axes of their elongations aligned each other and turned on, or if a slit of light positioned and oriented on the Off sub-area is turned off, the cells yield firings. If firings to the stimulations of the On sub-areas are plotted positively and those to the Off sub-areas negatively with respect to the positions of the slit, the profile of the firings is described with a Gabor function [2]: A sinusoid with a phase parameter between 0 and 2 π multiplied by a Gaussian. Interestingly, if many simple cells are sampled and each of them is best described by a Gabor function, their phases are distributed not merely around 0 and π /2, i.e. as those of cosine and sine Gabor functions, but uniformly between 0 and 2 π [3]. Based on this evidence, we have already proposed a neural model for simple receptive fields, where several sim- ple cells are arranged in a ring with mutual excitatory or inhibitory connections of fixed weights [4]; Phases of their receptive fields thereby varied gradually with rota- tion along the ring. 220 T. Hamada and K. Okada Meanwhile, such simple receptive fields are known to be basically organized before the eyes are naturally opened in kittens [5]. This suggests that the organizing process does not require any visual experiences. As a matter of fact, experimental studies suggested that the process could be due to some correlations in dark or spontaneous firings in the lateral geniculate nucleus (LGN) [6, 7]. A pioneering theoretical study [8] assumed the correlation to have the shape of "Mexican hat", i.e. the firings are correlated positively among nearby cells and negatively among farther cells. As- suming correlations of the "Mexican hat" as well as the ring structure, we have theo- retically shown that cortical cells in the ring, whose synaptic weights from LGN are initially set to be random, develop simple receptive fields with their phases gradually varied between 0 and 2 π along the ring [9]. This study, however, concerned only with one-dimensional profiles of receptive fields perpendicular to their preferred orienta- tions. We here propose a model, again with the ring structure, for the development of two-dimensional profiles of simple receptive fields. Computer simulation revealed that the cortical cells develop simple receptive fields whose phases gradually vary with rotation along the ring, only if the correlation was assumed to have a shape of Gaussian as recently shown experimentally in the developing LGN of ferrets [10], but fail to develop such receptive fields if the shape of “Mexican hat” is assumed for the correlation. 2 Methods 2.1 Structure of the Model For the sake of simplicity, we consider only a set of cortical cells whose receptive fields will be the same as to orientations, ocular dominances and retinal positions, but dif- ferent as to phases. We assume that several cortical cells Cp, p=1~n, are arranged in a ring (the upper part in Fig. 1a) with mutual intra-cortical connection Ip,q (p, q=1~n): ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ − × = n q p I q p π 2 cos , for q p ≠
=0.1 for q p = . (1)
The interaction is excitatory between nearby cells and inhibitory between diagonally opposed cells on the ring (see 2.4 for detail). The ring structure is meant to be merely functional. We consider two two-dimensional arrays of LGN cells, one for On center cells and the other for Off center cells, each with the size m x m (the lower part in Fig. 1a). Each of the On LGN cells yields an excitatory synapse to each of the cortical cells: The synapses from the On array to a cortical cell Cp (p=1,…, or n) thereby compose a two-dimensional m x m list of weights
. Similarly W off p for Off LGN cells.
A Ring Model for the Development of Simple Cells in the Visual Cortex 221
p,q and arrays of On and Off LGN cells. b: Correlated spontaneous activities among LGN cells of the same type (upper) and of the different types (lower) in one of the trials with correlations in (2) and (3). 2.2 Correlation of Spontaneous Activities in LGN Although correlations of spontaneous activities have been physiologically studied in adult retina of cats [11] and in developing LGN of ferrets [10], those in the developing LGN of cats are not yet studied physiologically. We thereby tested two types of func- tions for the correlations. The function firstly tested for correlations between cells of the same type, i.e. among On center cells or among Off center cells, is a Gaussian in (2) where r is the separation between the centers of their receptive fields and σ
the standard deviation of the Gaussian. Thus, spontaneous activities of two LGN cell of the same type are assumed to be more positively correlated if their receptive fields are closer to each other. ⎟⎟ ⎟
⎞ ⎜⎜ ⎜ ⎝ ⎛ × − = σ c r r C same 2 2 2 exp
) ( . (2) Correlation between spontaneous activities of LGN cells of different types, i.e. one On center cell and the other Off center cell, is assumed to follow the function (2) multiplied by -0.3: ) (
. 0 ) ( r r C C same diff × − = .
(3) Thus, LGN cells of different types are negatively and weakly correlated only if their receptive fields are close to each other.
222 T. Hamada and K. Okada The function secondly tested for correlations between LGN cells of the same type has the shape of "Mexican hat": ( ) ⎟⎟
⎠ ⎞ ⎜⎜ ⎜ ⎝ ⎛ × − × − ⎟⎟ ⎟ ⎠ ⎞ ⎜⎜ ⎜ ⎝ ⎛ × − = σ σ
c r r r C same 3 2 2 2 2 2 exp
1 . 0 2 exp
) ( . (4) Thus, spontaneous activities of two cell of the same type are positively correlated if r is small, but negatively at larger values of r. Correlation between cells of different types is assumed to be the function of (4) multiplied by -0.3. The developmental process is assumed to be an iteration of many (usually more than 5000) trials, in which a set of xy coordinates for a point as the center of correlations is randomly selected in either of the two LGN arrays and correlated activities spread around the point on the array according to the same-type function. Additionally, cor- related activities spread around a point with the same coordinates on the other array according to the different-type function. Fig. 1b explains such correlated spontaneous activities in one of the trials, where correlations in (2) and (3) are used. The point is incidentally selected at {x, y}={7, 6} on the On array and correlated activities is thereby positive on the On array and negative on the Off array. In another trial, the point would be differently positioned, correlation on one of the arrays would be incidentally positive and that on the other array would be negative. Correlated spontaneous activi- ties on the On and Off arrays in the i-th trial are named spont on i and
spont off i , respectively. 2.3 Initial Synaptic Weights At the earlier stage of development, LGN axons are attracted toward a cortical neuron presumably according to a gradient of diffusible molecules from the cortical cell [12]. Hence, we suppose that the closer to the center of the array a LGN cell is positioned, the stronger tends to be the synaptic weight from the LGN cell to one of the cortical cells. For describing this tendency, an arbor function A(r) is defined: ⎟⎟ ⎟
⎞ ⎜⎜ ⎜ ⎝ ⎛ × − = σ a r A r 2 2 2 exp
) ( . (5) It is a Gaussian with respect to distance r of a LGN cell from the center of the array, with the standard deviation σ
. The list of initial synaptic weights
=
(q=1~n) is thereby assumed to be a list (m x m) of random numbers between 0 and 1, Rand, multiplied by A(r):
A Ring Model for the Development of Simple Cells in the Visual Cortex 223 )
0 r A Rand on q W i × = = .
(6) Similarly for W off q i 0 = . Fig. 2a shows the W on q i 0 = for q=1~n in the case n=5. 2.4 Modification of the Synaptic Weights Activity of a cortical cell in a trial is determined by inputs from the LGN cells as well as from the other cortical cells in the ring. Inputs from LGN cells to a cortical cell Cq (q=1,…, or n) in the i-th trial, pre q i , is described by spontaneous activities on the On LGN array multiplied by
plus spontaneous activity on the Off array at the i-th trial multiplied by
:
spont W spont W pre off off q on on q q i i i i i × + × = . (7) The final activity of Cp in the i-th trial, post p i , is a sum of the inputs from the LGN cells plus inputs from the other cortical cells in the ring, which is described as:
× = ∑ = 1 , .
(8) Note that I q p , (p, q=1~n) is defined so that inputs from LGN are only weakly weighted by 0.1, while inputs from the cortical cells are weighted by 1*Cosine function (see the equation (1)). Such a design of I q p , is based on the anatomical evidence that the number of synaptic buttons that a layer 4 cell receives from LGN cells occupies only a few percent among the number of synapses on its cell body and dendrites [13], which means that synaptic influences from the other cortical cells are much stronger than those from LGN cells. Synapses in the visual cortex are known to be modifiable during development [12]. We thereby assume that the synapses are modified due to the Hebb rule: Synaptic weight is strengthened when the pre-synaptic activity is correlated with the postsy- naptic activity and weakened otherwise. Changes in W on p between i-th and i+1-th trials are then hypothesized:
224 T. Hamada and K. Okada post pre W W p i p r A on p on p i i i × × × + = + ) ( 1 λ . (9) where
λ is a constant which determines the rate of development. Similarly for W off p i 1 + . We assume that the total synaptic weights over each cortical cell are conserved: W on p i 1 + + W off p i 1 + = const . (10) Note that On and Off synaptic weights are conserved jointly [14], but not separately as in [10]. The formula in (9) is computed under the constraint of (10) in a trial, and the trials are iterated until the values converge. 2.5 Visual Responses After the synaptic weights are converged, visual responses of the cortical cells are computed. We assume the receptive field of a LGN cell positioned at
0 on the array has a Gaussian form, where σ
is the standard deviation of the distribution. ( ) ⎟⎟ ⎟ ⎠ ⎞ ⎜⎜ ⎜ ⎝ ⎛ − = × − σ 2 2 2 0 exp 0 r r r r R .
(11) A light spot at a position r 0 yields inputs from LGN to Cq, previs q r 0 , as: W R W R previs off q on q q r r r × − × = 0 0 0 . (12) The final visual activity in Cp due to the spot at r 0 , postvis p r 0 , results from I q p, :
previs postvis q p n q q p r r , 1 0 0 × = ∑ = . (13)
The visual receptive field of cell c p is a spatial summation of postvis p r 0 with respect of r 0 . A Ring Model for the Development of Simple Cells in the Visual Cortex 225
The parameters were set as follows, if not noted otherwise: n=5, m=15, σ
=3, σ
=0.5, σ
=1.5, λ
3 Results 3.1 Development of Simple Receptive Fields Each of the subfigures in Fig. 2a is the initial synaptic weight W on p for p=1,…or n which are random numbers multiplied by a Gaussian envelop. Note that although these subfigures, as well as those in the other columns in the figure, are vertically arranged in order of p, they actually constitute a ring structure. Fig. 2b shows the synaptic weights,
in the left column and W off p in the right, after 10000 iterations when the Gaussians in (2) and (3) are used as the correlation functions. Note firstly that each of
and
W off p (p=1, …, or n) is segregated, i.e. strengthened in some sub-areas and weakened in the other sub-areas. Secondly, the boundaries between the sub-areas
Fig. 2. a: Initial W on . b: W on (left) and W off (right) after trials with Gaussian correla- tions. c: Visual receptive fields, resulted from the synaptic weights in b and I p, q
. Bars are added for clarifying translations of the receptive fields along the column. d: Visual receptive fields, after trials with the Mexican hat correlations. White arrows are added for clarifying rotations of the fields along the column. The subfigures in each of the columns are actually arranged in a ring. 226 T. Hamada and K. Okada are roughly straight. Thirdly, configuration of the sub-areas is similar between two adjacent cells on the ring, but mostly reversed between diagonally opposed cells, so that the boundaries gradually translate with rotations along the ring: In the case of Fig. 2b, the boundaries gradually translate toward the lower-right along the downward direction in the column. Besides,
and
W off p for the same cell C p (p=1,…, or n), i.e. two subfigures in one of the rows of Fig. 2b, are reversed images each other: Sub-areas which are strongly innervated by synapses from On LGN cells are weakly innervated by synapses from Off LGN cells, and vice versa. Fig. 2c shows visual receptive fields, i.e. distribution of visual responses due to the synaptic weights in Fig. 2b as well as due to the intra-cortical interaction
, which
again show the properties described above for W on p and
W off p . As an example, the receptive fields are composed of On and Off sub-areas, each of which corresponds to the sub-areas in W on p and
W off p . Besides, the configurations of the sub-areas gradually translate with rotation along the ring.
Download 12.42 Mb. Do'stlaringiz bilan baham: |
ma'muriyatiga murojaat qiling