|Ph.D Student||Harel Yuval|
|Subject||Optimal Encoding of Dynamic Stimuli in Sensory Neural|
|Department||Department of Electrical Engineering||Supervisor||Professor Ron Meir|
|Full Thesis text|
Biological agents have access only to partial and noisy information about the state of their environment, as obtained through sensory organs and encoded in the firing patterns of sensory neurons. Since the sensory system is a product of natural selection, it may be possible to explain details of sensory encoding as the result of optimization with respect to some performance criterion corresponding to the goals of the acting organism. Such criteria may involve state estimation errors, energy expenditure, or motor control performance. Obtaining concrete predictions from this optimality hypothesis is challenging, since characterization of optimal encoding is intractable except in very simple models. In particular, the computation of performance criteria in specific models generally involves the posterior distribution of the state given the history of neural spikes, giving rise to a nonlinear filtering problem.
Estimation of the environment state from neural spike patterns is known as neural decoding. The neural decoding problem has been solved analytically only for homogeneous neural populations that uniformly cover the stimulus space. The optimal encoding problem poses additional difficulties, even in this simplified model. Many previous works considered easier proxies to the encoding problem, such as maximization of Fisher information or minimization of decoding error for non-optimal decoders. We present an approximate filter solving the decoding problem for dynamic multivariate stimuli and for more general sensory models. The closed-form decoder provides insight into the nature of optimal neural codes that is not easily gleaned from numerical methods.
We study optimal encoding both in the uniform case, where estimation error may be evaluated in closed form, and in the non-uniform case, where the approximate filter is used. Optimal codes are found to match experimentally observed neural codes. We show that when minimizing decoding error in a multivariate setting, additional constraints may be necessary to obtain a well-posed problem. Minimization under biologically plausible constraints yields solutions that qualitatively differ from those previously obtained by optimizing proxies to decoding error.