Latent variable decoding in biological and artificial agents

Abstract

"Decision-making in the presence of uncertainty is a pervasive computation. Latent variable decoding—inferring hidden causes underlying visible effects—is commonly observed in nature, and it is an unsolved challenge in modern machine learning. On many occasions, animals need to base their choices on uncertain evidence; for instance, when deciding whether to approach or avoid an obfuscated visual stimulus that could be either a prey or a predator. Yet, their strategies are, in general, poorly understood. In simple cases, these problems admit an optimal, explicit solution. However, in more complex real-life scenarios, it is difficult to determine the best possible behavior. The most common approach in modern machine learning relies on artificial neural networks—black boxes that map each input to an output. This input-output mapping depends on a large number of parameters, the weights of the synaptic connections, which are optimized during learning.(...)"