Vision: Building feature maps

The primary visual cortex (area V1), which receives projections from both eyes via the thalamus, is the first cortical region to process the visual world. Unlike its inputs, cells in V1 are selective to higher-order features, the most pronounced one being the orientation of an edge. Optical imaging reveals that orientation preferences vary smoothly along the cortical sheet thus forming an orientation map, so called because the cells' preferred locations in visual space vary smoothly as well.

Orientation maps self-organize early during cortical development— their formation does not require visual experience.

The origin of orientation maps is still unknown, however they appear in kittens reared in complete darkness—suggesting that intrinsic factors regulate map formation. One widely held hypothesis is that, starting from a tabula rasa, internally-generated activity patterns (in the retina and thalamus) orchestrate map formation through learning, the precise adjustment of synaptic strengths. However, it is unclear how this precision is achieved given that the underlying molecular mechanisms are stochastic.

Cortical feature maps could arise in their mature form through a pattern-formation process akin how to spotted patterns that form on a leopard's coat.

We explored the hypothesis that electrical pattern-formation can seed an orientation map with no need for learning. A two-dimensional recurrent network that can form bumps of activity achieved this. The same input yields repeatable bump configurations because bumps are seeded by innate neuronal heterogeneity. Differently oriented inputs select different bump configurations because they displace seeded bumps in different directions. Thus, this heterogeneous network satisfies the two requirements for orientation selectivity. Since this idea exploits variability, it is well-suited for building complex systems out of imprecise components.

Students
Kareem Zaghloul's silicon retina includes four ganglion-cell types.
Paul Merolla's visual cortex chip has orientation maps.

Collaborators
Leif Finkel
Matthew Dalva
Marcos Frank

Funding
ONR's MURI Program
Packard Foundation's Interdisciplinary Science Program
NSF-NIH's CRCNS Program