Retinomorphic vision systems : reverse engineering the vertebrate retina
This thesis seeks to explain how the retina satisfies both top-down constraints (functional) and the bottom-up constraints (structural) by analyzing simple physical models of the retina and mimicking its structure and function in silicon. In particular, I examine spatiotemporal filtering in the oute...
| Summary: | This thesis seeks to explain how the retina satisfies both top-down constraints (functional) and the bottom-up constraints (structural) by analyzing simple physical models of the retina and mimicking its structure and function in silicon. In particular, I examine spatiotemporal filtering in the outer plexiform layer of the vertebrate retina, and show how outer retina processing is augmented by further processing in the inner plexiform layer, creating an efficient implementation that encodes moving stimuli efficiently over a wide range of speeds.
My working hypothesis is that biological sensory systems seek to optimize both functional and structural constraints. On the functional side, they must maximize information uptake from the environment while they minimize redundancy in their outputs. On the structural side, they must maximize resolving power in space and time, by making the processing elements small and fast, while they minimize wiring and energy consumption. If structure and function did indeed coevolve, as I assume, studying how structural and functional constraints are optimized simultaneously is our only hope of understanding why nature picks the solutions that we observe.
Addressing both structural and functional constraints requires combining science and engineering. Scientists study an existing structure, and seek to understand how it functions in an optimal or near-optimal fashion, based on theoretical grounds. Rarely does a scientist ask: Will the structure be more cost effective, more reliable, or more reproducible if a less-than-optimum function is chosen? Engineers, on the other hand, design an optimal implementation for some desired function, based on an existing set of standard primitives. Rarely does an engineer ask: Is this the most natural set of primitives to use for this particular function? Thus, neither discipline attempts to optimize both function and structure globally. In contrast, evolution, operating in a purely opportunistic fashion, continuously seeks increasingly elegant solutions that meet these constraints.
For these reasons, I have adopted a multidisciplinary engineering-science approach that combines analysis with synthesis. When tailored synergestically, this approach can shed light on questions about which neurobiologists care, while advancing the state of the art in sensory-system design. |
|---|