KERNEL REGRESSION FOR NEURAL SYSTEMS IDENTIFICATION

Maneesh Sahani

Gatsby Computational Neuroscience Unit

University College, London

 

 

Techniques from the field of systems identification have been very productive in experimental neuroscience as a means of characterizing either the encoding transfer functional of a neuron F[I(t)] =s(t) (Marmarelis and Marmarelis, 1978), or the associated decoding functional G[s(t)] =I(t) (Rieke et al., 1997); here I(t) is the sensory input, while s(t) is the spike train.  The simplest form of systems identification involves estimation of the coefficients in the Volterra series expansion of the desired functional.  However, due to data limitations and the high dimensionality of higher-order coefficients, estimation of more than the first two terms has proven intractable.  With discretized signals, it is easily seen that the estimation of the Volterra coefficients is equivalent to linear regression of the system output on non-linear functions of lag-vectors of the input (note “system output” and “input” may be reversed with respect to the neuron, depending on whether we are expanding the encoding or decoding functional).  In the first part of this talk, I will discuss the use of techniques form the linear regression literature, specifically ridge-regression and automatic relevance determination (ARD), to stabilize estimates of higher-order Volterra coefficients in the face of limited data.   In the second part, I will argue that regression using the conventional Volterra series expansion is equivalent to kernel regression (in the sense of support-vector or relevance-vector regression) using a series of polynomial kernels.  This viewpoint allows for a relatively compact representation of the coefficients in terms of data vectors, and, in combination with ARD, allows estimation of still higher-order Volterra coefficients.  Furthermore, regression using kernels other than polynomial, leads to non-linear expansions of the transfer functional different from the Volterra series.  These alternative expansions may provide additional insight into the neural code.