Innate visual learning through spontaneous activity patterns

Abstract

Patterns of spontaneous activity in the developing retina, LGN, and cortex are necessary for the proper development of visual cortex. With these patterns intact, the primary visual cortices of many newborn animals develop properties similar to those of the adult cortex but without the training benefit of visual experience. Previous models have demonstrated how V1 responses can be initialized through mechanisms specific to development and prior to visual experience, such as using axonal guidance cues or relying on simple, pairwise correlations on spontaneous activity with additional developmental constraints. We argue that these spontaneous patterns may be better understood as part of an "innate learning" strategy, which learns similarly on activity both before and during visual experience. With an abstraction of spontaneous activity models, we show how the visual system may be able to bootstrap an efficient code for its natural environment prior to external visual experience, and we continue the same refinement strategy upon natural experience. The patterns are generated through simple, local interactions and contain the same relevant statistical properties of retinal waves and hypothesized waves in the LGN and V1. An efficient encoding of these patterns resembles a sparse coding of natural images by producing neurons with localized, oriented, bandpass structure-the same code found in early visual cortical cells. We address the relevance of higher-order statistical properties of spontaneous activity, how this relates to a system that may adapt similarly on activity prior to and during natural experience, and how these concepts ultimately relate to an efficient coding of our natural world.

Figure 1. Experimental and theoretical 2D spontaneous activity images.

(A) Experimental wave propagation: calcium imaging of a retinal wave (data as described in [9]). (B) Physiological model wave propagation: the ganglion cell layer activation of a retinal wave model (data from model described in [17]). (C) Physiological model wave extent: simulated retinal wave propagated to fullest extent (adapted from [18]). (D) Abstract model wave extent: a pattern generated by the technique used in this paper with parameters (p = 0.55, r = 3, t = 6) as detailed in the methods section.

Figure 2. V1 simple cell receptive fields derived through an efficient coding of natural scenes and spontaneous activity patterns.

(A) Receptive fields from sparse coding: basis functions derived from natural images (algorithm as described in [40]). (B) Receptive fields from ICA: filters derived from natural images (algorithm as described in [48]). For panels (B,C,D) the same patch collection and efficient coding algorithm was used as detailed in the methods section. (C) Receptive field filters derived from images of simulated retinal waves as in figure 2 of - a few examples are in figure 1c of this paper. Patch size for this data corresponds to approximately 0.3 mm. Refer to the text for the implication of this result. (D) Receptive field filters derived from our generated patterns with parameters (p = 0.7, r = 3, t = 8).

Figure 3. Summary of pattern sampling and analysis demonstrating relevant variation in derived gabor filters.

(A) Phase plane with ‘r’ fixed at 3. The curved line in the plot indicates the phase transition boundary as detailed in the methods section. The transparent color contours below the phase transition line indicate the trend for the median orientation bandwidth in that area of the plane. (B) Sampled patterns from (p,r,t) space near the critical percolation threshold - (0.15,3,1), (0.48,3,5) and (0.83,3,10). (C) The 16×16 pixel derived ICA filters. (D) Seven parameter gabor fits of those filters. (E) Histograms of the gabor orientation bandwidths in blue compared to the physiological median in red.

Figure 4. Information relevant to filter formation goes beyond simple, pairwise correlations.

(A) Examples of training data for efficient coding. Whitened images were obtained by flattening an assumed 1/f slope in the Fourier amplitude spectrum. (B) PCA bases from each of the six data sets in (A). Note, algorithms which rely on pairwise correlations alone (also known as second order image statistics) only find structure for receptive field formation in correlated data [red rectangle] although much of the useful structure still visible in the decorrelated (“whitened”) images is not captured. Also note the receptive fields in this case are not localized. (C) Representative filters from the same image sets using ICA. Note that the filters from the whitened and unwhitened, natural and wavefront-patterned data qualitatively resemble receptive fields [red square], whereas unstructured 1/f noise does not produce equivalent filters - unlike the results of pairwise correlation-based measures. (note: differences in whitened ICA filter sizes are primarily a product of the 1/f assumption).