Maneesh Sahani

Computation and Neural Systems Program, Caltech

I describe a method for automated characterization of response-functions of sensory neurons. The technique involves online feedback-directed exploration of the neuron's response using a modified active-learning algorithm. It can be applied to learn the stimulus-dependent functional form of any scalar measure (such as firing rate, latency or correlation) of the response of one or more neurons,

Sensory neurons, particularly those deeper than the first few layers in a sensory pathway, often represent high-dimensional stimulus spaces in a non-linear fashion. A number of methods have been used to characterize these responses. The most common is still trial and error, where the experimenter chooses a small, discrete set of stimuli (sometimes a basis spanning a relevant space) and looks for variation in firing rate. More systematic methods have involved characterization of either forward or reverse transfer functions by white-noise analysis, (e.g. Bialek et al. 1991), and systematic search for maxima (e.g. Tzanakou et al. 1979, Nelken et al. 1994). These have drawbacks: the white-noise approach uses the entire stimulus space, much of which the neuron may not respond to. It is most successful on linear transfer functions. The maxima search techniques that have been applied are Markovian (alopex), or close to it (simplex), and so inadequately exploit data already gathered. Both techniques, when successful, return a single point in stimulus space - the mean or a local maximum of the response-function - when we may be more interested in the shape of the function.

I take the view that the problem to be solved is one of noisy function approximation, and adapt techniques from the machine learning literature to the task. At each step, an estimate of the response-function is computed and used to direct further sampling of the response, so as to minimize uncertainty in regions of high activity. Thus, in contrast to the white-noise approach, the space is explored non-uniformly, with regions of interest being emphasized. All data gathered up to some time are used to derive the estimate of the response-function at that time, and therefore to direct subsequent samples. The result of the process is a function that approximates the response. Samples are directed so as to make this approximation most accurate in regions of high activity.

Wed Mar 12 22:07:02 PST 1997