Venkatesh Murthy (Harvard University)
Abstract. In a cluttered sensory environment, animals must segregate objects of interest from ever varying backgrounds, and categorize them for decision making. We study this problem in the olfactory system. We have shown that laboratory mice are excellent at recognizing individual odorants embedded in unpredictable and variable background odor mixtures. To understand the limits of this ability and the potential algorithms involved, we combine behavioral and neural measurements in mice, with computational analysis. Using the earliest sensory representation of odors (olfactory receptor activity) as input, we find that linear decoders with sparse weights can match the behavioral performance of mice. Interestingly, both linear decoders and mice perform poorly if they are trained with single odors and then tested with odor mixtures. These results may have some implications for discriminative vs. generative models in learning.