Decoding in neural systems: stimulus reconstruction from nonlinear encoding

Abstract — The encoding of information about the outside world in the temporal activity of sensory neurons is an extremely complex process that has eluded the understanding of the scien-tific community for decades. The reconstruction of sensory stim-uli from observed neuronal activity provides a basis within which we might ascertain the nature of the sensory information encoded by the cells. We present a decoding strategy for predicting the sensory stimulus from the neuronal response that is based on the mechanisms of encoding. For a class of encoding mechanisms characterized by a linear function followed by a memoryless non-linearity, referred to as Wiener systems, the Bayesian estimator is derived from the transformational properties of the nonlinear-ity. The result is a reconstruction paradigm in which the ability to predict sensory stimuli from the neuronal response depends heavily upon how well the encoding process has been character-ized, and thus provides a measure of our understanding of the underlying physiological process