|M.Sc Thesis||Department of Electrical Engineering|
|Supervisor:||Prof. Meir Ron|
|Full Thesis text|
Sensory systems are constantly presented with the challenge of inferring the state of the environment based on noisy and partial measurements. Using ideas from the theory of optimal filtering, considerable progress has been achieved in the field of neural decoding by modeling the neural responses of sensory cells as point processes, with an instantaneous input-dependent rate. However, despite biological evidence showing that sensory neurons' responses often depend on the history of their input signal, most approaches to neural decoding assume an instantaneous response profile, ignoring much of the rich temporal structure in the input.
Abstract In this work, we apply a continuous time, continuous state-space nonlinear filtering approach to the problem of estimating an analog vector signal, based on measurements in the form of a point process (a.k.a. spike train). The rate, or intensity, of this process is a function of the spatio-temporal sensory input signal, and is implemented through a spatio-temporal filter. Consistently with recent experimental findings, different spatio-temporal filters are used, emphasizing different aspects of the input signals. This approach allows us to broaden the scope of previous decoding schemes by expanding the family of response functions encompassed in the model, thereby increasing the model's biological plausibility. In addition, we examine the effect of encoding by use of multiple input signal features, resulting from the application of different temporal filters. By adding Kalman filtering to our decoding scheme we show that the additional information in the input, conveyed by this encoding scheme, can be exploited to achieve better performance. Based on known spatio-temporal filtering properties of retinal ganglion cells, we demonstrate the performance of our decoding scheme on several types of input signals.