Generalization from one exemplar in mice and neurons

Learning general principles from single examples is a hallmark of high cognitive functions. This type of learning is computationally challenging for artificial agents, yet many animals such as primates can perform it with ease. To study one-exemplar generalization in the lab, it would be ideal to develop appropriate behavioral tasks in more tractable model organisms, such as rodents. Here we report that mice can perform one-exemplar generalization with difficult texture-based visual stimuli such as “leaves” and “rocks”. We trained mice to discriminate between single images from two texture categories, using water rewards for one of the images. Once the mice were trained, we introduced two unrewarded test images from the same two texture categories. Mice responded to the new image from the rewarded category, despite the lack of reinforcement, but not to the new image from the unrewarded category. The same task could not be solved by low-level features from an artificial neural network. To test what neural properties may give rise to generalization, we recorded simultaneously from large neural populations in primary and higher-order visual cortex in naive mice, while presenting the same set of images we used for the trained mice. Simple linear decoders trained on neural responses to single exemplars were able to generalize to new exemplars, similar to the behavior of the mice. To look for changes in neural coding after learning, we also recorded large neural populations in trained mice, which we are currently investigating. Our results establish mice as a viable model to study generalization from one example and thus enable a wide range of neuroscience techniques to be used for identifying the generalization algorithms used by the brain.