A silicon model of the primary visual cortex: Representing features through stochastic variations

We present the bump chip, a silicon model of the primary visual cortex that proposes a design principle for building orientation maps. The origin of orientation maps is still unknown, however, recent evidence shows that they appear without the need for visual experience, and are remarkably robust to experimental manipulations. It is widely thought that internally-generated activity patterns, which drive the cortex via afferent inputs, orchestrate map formation through Hebbian learning. The bump chip, on the other hand, obtains its selectivity through a recurrent network that forms patterns of neural activity spontaneously; these patterns, which are seeded by random component mismatch, serve as the scaffold of the map. Therefore, our chip predicts that disrupting afferent activity during cortical development will not alter the layout of orientation selectivity. The design principle used by the bump chip can help engineers build complex systems with imprecise components. Our chip exploits component variability to obtain two traits that set it apart from every other man-made system to date: (1) its functional architecture (orientation selectivity) is not specified prior to fabrication, and (2) the architecture\u27s scaffold is innate to each chip and exists as an indelible imprint. Therefore, our chip attains all of the benefits of a self-organizing learning system without having to go through the tedious process of learning