Applications of Computational Analyses to Understand Visual Development

In recent years, there has been an increase in application of advanced mathematical and computational techniques to study visual encoding during development. Mark Albert, Assistant Professor from the Department of Science at Loyola University, introduces the innate learning strategy to overcome limitations of known models of visual development. Current visual models dichotomize visual experience as either being innate (visual properties found at birth) or learned (properties dependent on experience). In his paper, “Innate Visual Learning Through Spontaneous Activity,” Albert uses a computational model to show how highly “structured, spontaneous endogenous activity” guides the development of the visual system prior to birth (Albert 2008, pg. 1). According to Albert, this activity is exhibited in the retina, thalamus, and visual cortex prior to birth. In the thalamocortical visual pathway (retina -->LGN<--> V1), this neuronal activity is necessary to mediate dendritic patterning, axonal branching, and synaptic activity to aid in early visual development.

In his lecture, he elaborated the role of V1 simple cells, which are stimulated by bright and dark regions of a natural image, in refining of natural scenes. He used the innate learning strategy to explain how newborn animals can share the V1 response properties of the adult visual cortex before any visual experience is formed. Using the developmental timeline of a ferret, he elaborated on the visual activity exhibited prior to the eye opening. From birth to day 25, retinal waves are detected, and by day 28, spontaneous activity is seen in the LGN. The LGN/V1 spontaneous activity has more direct impacts on cortical receptive field formation. During the talk, he mentioned how using higher order statistical models, such as Ising models, can help our understanding of the relations between spontaneous activity and V1 development. He explained how the efficient coding hypothesis argues that the V1 simple cells in adult mammals are responsible for fine tuning and encoding of the visual world by removing statistical redundancy.

From a computational perspective, many models on visual encoding are limited in their statistical approaches to simple pair wise correlations, and argued for the need to apply advanced statistics to models. In his paper, he generated spontaneous and abstract patterns through local interactions and found that these patterns resembled retinal waves. Albert used higher-order statistics to bridge the gap between models of sparse efficient coding—a higher level and more relevant model—and spontaneous activity. In sparse coding, computations are done with the least amount of active neurons. With this coding, he showed that the same algorithm used for encoding of natural input in V1 cells in adults can be applied to prenatal spontaneous visual activity. He used a percolation model, because it was analogous to the retinal wave model to generate patterns, and used the encoding algorithm to show that they produce the 2D Gabor code found in V1 cells. Ultimately, this production is significant because it shows that endogenous spontaneous activity trains the early visual system for a coding of the world, and abolishes the nature (innate) versus nurture (experience) dichotomy (Albert 2008, pg. 7). These patterns are sustained prior to birth and adapt through natural experience.

            The use of sparsity as a computational technique can extend to other areas of visual development. For instance, it can address variations in responses between normal and abnormal visual input that have yet to be studied computationally. Use of efficient sparse coding allows for a reproduction of models that are consistent with established data. In “Sparse Coding Can Predict Primary Visual Cortex Receptive Field Changes Induced by Abnormal Visual Input,” Hunt, Dayan, and Goodhill (2013), contend that the sparse coding hypothesis accounts for the outcome of cortical plasticity seen in visual development. The visual experience has profound impacts on neural responses. For example, neural activity to stripes has been shown to orientation specific, a response solicited by the visual experience. Under abnormal rearing conditions, sparsity has been detected in receptive field acquisition. Sparse coding models, thus far, have only analyzed the relationship between normal visual stimuli and development of monocular receptive fields. In their study, they applied sparse coding models to binocular receptive field development across six abnormal rearing conditions (normal, stripe reading, orthogonal stripe rearing, monocular deprivation, random monocular deprivation, monocular occlusion). By modifying the visual experience for cats, they tested the power of sparsity to see whether it can be used to understand the development of V1 simple cell responses. Under all conditions, changes in the receptive fields correlated with established experimental values.  Therefore, sparse coding can be used to simulate the structure of V1 receptive fields that arise when animals are reared under both normal and abnormal visual input (Hunt et. al, 2013, pg. 8). Ultimately, sparse coding has the capacity to unify principles of visual development previously thought to be separate, be It spontaneous activity or receptive field modeling.

References

Albert MV, Schnabel A, Field DJ (2008) Innate Visual Learning through Spontaneous Activity       

Patterns. PLoS Comput Biol 4(8): e1000137. doi:10.1371/journal.pcbi.1000137

Hunt JJ, Dayan P, Goodhill GJ (2013) Sparse Coding Can Predict Primary Visual Cortex

Receptive Field Changes Induced by Abnormal Visual Input. PLoS Comput Biol 9(5):

e1003005. doi:10.1371/journal.pcbi.1003005