|M.Sc Student||Yaeli Eran Steve|
|Subject||Optimal Neural Encoding - Mathematical Foundations and|
|Department||Department of Electrical Engineering||Supervisor||Professor Ron Meir|
|Full Thesis text|
In naturally evolved organisms stimuli indicating the state of the external environment are encoded as neural spike trains. Arguably, these spike trains are later decoded by the central nervous system in order to estimate the environmental state, thereby guiding the organism's decisions and behavior. Although the theory of optimal neural decoding is well established, there are still many question marks concerning optimal neural encoding, and more specifically optimal neuronal tuning curves. This issue has been previously studied using lower bounds on the minimal mean squared estimation error (MMSE), which are based on Fisher information, but in many cases the predictions of such approaches contradict results based on the MMSE. Surprisingly, very little attention has been paid to optimality in the MMSE-sense, and no links to experimental results have been established.
In this work we examine optimal tuning curves directly within the Bayesian setting, using the common MMSE loss function. In simple cases the MMSE is derived analytically, whereas in more complicated cases that were not previously studied it is obtained through simulations. We find that optimal tuning curves are adaptive to the statistical nature of the environment, in agreement with ecological theories of sensory processing, and with experimental evidence. Moreover, we show that predictions based on lower bounds are often qualitatively false. Some of our results provide a theoretical justification for observed neurobiological phenomena, suggesting that neural systems may indeed dynamically adapt themselves so as to optimize their performance.