Neural modeling of efficient coding in primate mt and mstd

Abstract

I conducted two studies that explore the implications of the efficient coding hypothesis on neural mechanisms in primate MT and MSTd. The efficient coding hypothesis holds that neuronal populations, particularly those in perceptual areas, have sparse, hierarchical coding schemes that efficiently encode stimuli. In the first study, I introduced a self-tuning mechanism for capturing rapid adaptation to changing visual stimuli within a population of neurons. Model MT speed cell tuning curve parameters were continually updated to optimally encode a time-varying distribution of recently detected stimulus values. In two simulation experiments, I found this dynamic tuning yielded more accurate, lower latency heading estimates from downstream model MSTd cells compared to a static tuning. The second study investigated how simulated, complex stimuli affected model cell sensitivities derived via an efficient coding framework. Model cell sensitivities were derived using nonnegative matrix factorization (NMF) of an MT-like population coding of optic flow speed and direction. A previous study (Beyeler et al., 2016) found that basis vectors derived in this manner matched the results from neurophysiological studies of area MSTd in macaques. This prior work used simple, planar dot-environments which lacked a varied environmental structure and thus many visual features that affect the flow field generated by self-motion. I replicated portions of this prior work, then used the same methods and analyses to generate and evaluate the results for alternative stimuli from more naturalistic settings with higher spatial resolution. I found that model units derived with NMF used a coding schema similar to that found in primate MSTd, confirming the original result with more complex stimuli.