Robustness and generality of orientation and ocular-dominance map formation in the BCM lateral-inhibition network trained on natural stereo images (Shouval–Intrator–Cooper, 1997)

A 1997 Vision Research study by Shouval, Intrator and Cooper showed that a 24×24 sheet of BCM (Bienenstock–Cooper–Munro) neurons coupled by difference-of-Gaussians lateral inhibition, trained on 24 DOG-preprocessed natural images presented to partially overlapping left- and right-eye patches, concurrently develops orientation-selective receptive fields and ocular-dominance columns, with sharply peaked (kurtotic, sparse-looking) activity histograms, while a linear principal-component-analysis (PCA)/Oja network remains binocular. The authors acknowledged that the demonstration rests on single "typical" runs, essentially two lateral-kernel settings, a fixed preprocessing pipeline, and a simplified implementation in which lateral inhibition affects only cell activity rather than the full synaptic gradient, and they suggested but did not carry out a quantitative comparison with explicit sparse-coding methods. The sparse-code comparison thread was substantially addressed by Blais et al. (1998) at the single-cell level and by Willmore et al. (2012), who showed that plain BCM population codes are highly redundant and that adding contrast normalization restores an efficient, sparse, overcomplete code; Castellani et al. (1999) supplied fixed-point theory for laterally interacting BCM networks. No published work systematically quantifies how the emergent maps depend on the lateral-interaction parameters and kernel form, how sensitive they are to preprocessing and image ensemble, or what the activity-only simplification costs relative to the full gradient. The robustness-and-generality question therefore remains open as a BCM-specific modeling problem of mainly model-refinement and historical significance.

Background

Two classical observables of primary visual cortex (V1) development are orientation selectivity — a neuron's preference for the orientation of edges in its receptive field, the region of visual space from which it receives input — and ocular dominance, the relative strength of synaptic drive from the two eyes. Across the cortical sheet these properties are spatially organized into orientation preference maps and ocular dominance columns. The Bienenstock–Cooper–Munro (BCM) rule (Bienenstock, Cooper & Munro, 1982) is a synaptic-modification rule in which each synapse changes in proportion to presynaptic activity times a nonlinear function of postsynaptic activity, $\phi(c,\Theta_{M})=c\,(c-\Theta_{M})$, where the modification threshold $\Theta_{M}=E[c^{2}]$ is a sliding average of the squared postsynaptic activity: consistently weak activity weakens synapses and strong activity strengthens them. Because the threshold makes the rule sensitive to higher-order (not merely pairwise) statistics of the input, BCM can develop selectivity that linear Hebbian rules cannot; in particular, linear principal-component-analysis (PCA)-type rules — such as Oja's rule, a normalized Hebbian update that converges to the leading eigenvector of the input covariance — obey a two-eye parity invariance that prevents them from breaking binocular symmetry to generate ocular dominance. A separate line of work showed that learning rules which explicitly optimize response sparseness — neurons silent most of the time with occasional large responses, a distribution property measured by kurtosis (the normalized fourth moment, large for long-tailed, sharply peaked distributions) — develop oriented, localized receptive fields from natural images (Olshausen & Field, 1996, 1997; Földiák, 1990).

Building on a single-neuron precursor (Shouval, Intrator, Law & Cooper, 1996), Shouval, Intrator & Cooper (1997, Vision Research 37(23):3339–3342) asked whether BCM plasticity could organize a cortical network, not just single cells, in a realistic two-eye visual environment. They trained a 24×24 sheet of BCM neurons whose activity is $c(x)=\sigma\bigl(\sum_{x'} I(x-x')\,\sigma(m(x')\cdot d)\bigr)$ — $\sigma$ a sigmoidal transfer function, $m(x')$ the feedforward weight vector of the neuron at cortical position $x'$, $d$ the input pattern, and $I$ a lateral interaction kernel implemented as a balanced difference-of-Gaussians (DOG) with excitatory length scale $\sigma_{E}$ and inhibitory length scale $\sigma_{I}$. Inputs were 24 digitized natural images (256×256 pixels, man-made objects excluded) preprocessed by a DOG filter (center $\sigma_{1}=1$, surround $\sigma_{2}=3$ pixels) and presented as small, partially overlapping circular patches to the two eyes (receptive-field diameter 20 pixels, inter-eye shift $s$). The network concurrently developed orientation-selective receptive fields, graded ocular dominance, ocular-dominance bands, and orientation-map structure; activity histograms were sharply peaked near zero with heavy positive tails; and control networks trained with a linear PCA/Oja rule remained binocular. The authors report this as, to their knowledge, the first network model trained on natural images that develops ocular dominance and orientation selectivity concurrently.

The paper, however, presents single "typical" runs with essentially two lateral-kernel settings (for example $\sigma_{E}=1,\sigma_{I}=5$ versus $\sigma_{E}=0.7,\sigma_{I}=2$, the latter giving less continuous maps), a fixed preprocessing pipeline, no seed-to-seed variability analysis, and a deliberately simplified implementation: as the authors write, "we have used a simplified version of the full gradient in which the lateral inhibition affects only the cell activity" — the lateral kernel enters the activity $c(x)$ but not the synaptic-gradient update. The sparse/kurtotic character of the activity was shown only through histograms, without numerical sparseness metrics or comparison against explicit sparse-coding methods.

Problem Statement

Determine, quantitatively, the robustness and generality of the Shouval–Intrator–Cooper (1997) result that a two-dimensional lateral-inhibition network of BCM neurons, trained on DOG-preprocessed natural stereo images with inter-eye misalignment, concurrently develops orientation-selective receptive fields and ocular-dominance columns. The single scientific objective is to establish which features of the emergent organization are generic consequences of BCM plasticity acting on natural binocular image statistics, and which are artifacts of the specific simulation choices; the five numbered items below are sub-questions of that one objective, each isolating one axis along which the 1997 demonstration could fail to generalize. (1) Map dependence on lateral interactions: systematically map the emergent map morphology over the lateral-interaction parameter space $(\sigma_{E},\sigma_{I})$ and across alternative interaction-kernel forms, using quantitative map metrics (for example ocular-dominance band spacing and continuity, and orientation-map continuity together with pinwheel statistics, pinwheels being the singular points of the orientation-preference map around which the preferred orientation rotates through half a turn), and identify which quantitative parameter ranges reproduce specific experimentally observed map features. (2) Sensitivity to the input representation: quantify how the receptive fields and maps change with DOG preprocessing parameters, image-ensemble statistics and size (including versus the original 24-image set and versus ensembles containing man-made objects), receptive-field size, inter-eye shift, and alternative or multiscale retinal or lateral-geniculate-nucleus (LGN, the thalamic relay between the retina and visual cortex) preprocessing. (3) Cost of the simplified implementation: quantify how the receptive fields and maps produced by the implemented activity-only lateral inhibition (the kernel entering the activity $c(x)$ but not the synaptic-gradient update) differ from those of the network in which lateral interactions enter the full gradient. (4) Unmodeled biological factors: determine which conclusions survive when factors the 1997 model omitted — innate orientation biases, three-dimensional cortical anatomy, cortical-area boundary shapes, and the specifics of imaging or readout methods — are incorporated or controlled for within this framework; here 'incorporated or controlled for' means an explicit within-simulation control — a stated modification of, or added term in, the same 2D BCM network rather than a substitute model — and a conclusion counts as surviving when the map metrics of (1)–(3) are unchanged under that control. (5) Code comparison: quantify the emergent representation (sparseness, kurtosis, information content, population coverage) and compare it with explicit sparsity- or kurtosis-maximizing methods (Olshausen–Field sparse coding; Földiák-type anti-Hebbian models).

References

1. Shouval, H., Intrator, N., & Cooper, L. N. (1997). BCM network develops orientation selectivity and ocular dominance in natural scene environment. Vision Research 37(23), 3339-3342. DOI: 10.1016/s0042-6989(97)00087-4. https://doi.org/10.1016/s0042-6989(97)00087-4
2. Bienenstock, E. L., Cooper, L. N., & Munro, P. W. (1982). Theory for the development of neuron selectivity: orientation specificity and binocular interaction in visual cortex. The Journal of Neuroscience 2(1), 32-48. DOI: 10.1523/JNEUROSCI.02-01-00032.1982. https://doi.org/10.1523/JNEUROSCI.02-01-00032.1982
3. Shouval, H., Intrator, N., Law, C. C., & Cooper, L. N. (1996). Effect of binocular cortical misalignment on ocular dominance and orientation selectivity. Neural Computation 8(5), 1021-1040. DOI: 10.1162/neco.1996.8.5.1021. https://doi.org/10.1162/neco.1996.8.5.1021
4. Blais, B. S., Intrator, N., Shouval, H. Z., & Cooper, L. N. (1998). Receptive field formation in natural scene environments. Comparison of single-cell learning rules. Neural Computation 10(7), 1797-1813. DOI: 10.1162/089976698300017142. https://doi.org/10.1162/089976698300017142
5. Castellani, G. C., Intrator, N., Shouval, H., & Cooper, L. N. (1999). Solutions of the BCM learning rule in a network of lateral interacting nonlinear neurons. Network: Computation in Neural Systems 10(2), 111-121. DOI: 10.1088/0954-898X/10/2/001. https://doi.org/10.1088/0954-898X/10/2/001
6. Remondini, D., Castellani, G. C., Bazzani, A., Campanini, R., & Bersani, F. (1999). Development of selectivity maps in a BCM network using various connectivity schemes. In: Neural Nets WIRN Vietri-99 (Perspectives in Neural Computing), 302-307. DOI: 10.1007/978-1-4471-0877-1_35. https://doi.org/10.1007/978-1-4471-0877-1_35
7. Shouval, H. Z., Goldberg, D. H., Jones, J. P., Beckerman, M., & Cooper, L. N. (2000). Structured long-range connections can provide a scaffold for orientation maps. The Journal of Neuroscience 20(3), 1119-1128. DOI: 10.1523/JNEUROSCI.20-03-01119.2000. https://doi.org/10.1523/JNEUROSCI.20-03-01119.2000
8. Lee, A. B., Blais, B., Shouval, H. Z., & Cooper, L. N. (2000). Statistics of lateral geniculate nucleus (LGN) activity determine the segregation of ON/OFF subfields for simple cells in visual cortex. Proceedings of the National Academy of Sciences 97(23), 12875-12879. DOI: 10.1073/PNAS.97.23.12875. https://doi.org/10.1073/PNAS.97.23.12875
9. Cooper, L. N., Intrator, N., Blais, B. S., & Shouval, H. Z. (2004). Theory of Cortical Plasticity. World Scientific, Singapore. DOI: 10.1142/5462. https://doi.org/10.1142/5462
10. Willmore, B. D. B., Bulstrode, H., & Tolhurst, D. J. (2012). Contrast normalization contributes to a biologically-plausible model of receptive-field development in primary visual cortex (V1). Vision Research 54, 49-60. DOI: 10.1016/j.visres.2011.12.008. https://doi.org/10.1016/j.visres.2011.12.008
11. Olshausen, B. A., & Field, D. J. (1996). Emergence of simple-cell receptive field properties by learning a sparse code for natural images. Nature 381, 607-609. DOI: 10.1038/381607a0. https://doi.org/10.1038/381607a0
12. Olshausen, B. A., & Field, D. J. (1997). Sparse coding with an overcomplete basis set: A strategy employed by V1? Vision Research 37(23), 3311-3325. DOI: 10.1016/s0042-6989(97)00169-7. https://doi.org/10.1016/s0042-6989(97)00169-7
13. Földiák, P. (1990). Forming sparse representations by local anti-Hebbian learning. Biological Cybernetics 64, 165-170. DOI: 10.1007/BF02331346. https://doi.org/10.1007/BF02331346
14. Udeigwe, L. C., Munro, P. W., & Ermentrout, G. B. (2017). Emergent dynamical properties of the BCM learning rule. The Journal of Mathematical Neuroscience 7. DOI: 10.1186/s13408-017-0044-6. https://doi.org/10.1186/s13408-017-0044-6
15. Miikkulainen, R., Bednar, J. A., Choe, Y., & Sirosh, J. (2005). Computational Maps in the Visual Cortex. Springer, New York. DOI: 10.1007/0-387-28806-6. https://doi.org/10.1007/0-387-28806-6