Decoding Visual Input based on V1 Neuronal Activities with Particle Filtering

In this study, we investigated the use of particle filtering in reconstructing time-varying input visual signals based on Macaque V1 neurons’ responses. A multitude of hypothesis particles are proposed for reconstructing the input stimulus up to time t. A prediction kernel (consisting of the first- and second-order forward Wiener kernels, derived by regression) is used to predict the neural response at time t based on the estimated input signals in the 200 ms prior to t. The fitness of this estimated response in predicting the measured response at time t is used to weigh the importance of the various hypotheses. The hypothesis particle space is collapsed by re-sampling over time. We find this method quite successful in reconstructing the input stimulus for 30 out of 33 V1 neurons measured. It out performs the optimal linear decoder that we have experimented with in the past