On sample eciency in artifical and biological agents

Abstract

Many of the predominant learning paradigms in modern machine learning techniques are inspired by biology. Deep networks are inspired by the neurons in the brain, and reinforcement learning was originally inspired by operant conditioning in cognitive science. However, there are still signicant gaps between artifical and biological models of intelligence. One notable difference is sample eciency - animals, in particular humans, are able to understand concepts and solve tasks with only a few demonstrations, whereas artifical agents take millions of examples and make just as many mistakes. This paper aims to quantify the raw difference in sample efficiency, accounting for the signicant priors held by biological agents. We do this by designing a suite of tasks that must be solved by both agents, in which each task requires a significantly dierent aspect of general intelligence to be solved. We train artifical and biological agents on said tasks after initial learning stages to control for significant priors, and finally examine and compare the quality and efficiency of learning in these tasks, such that we might understand the relationship between the sample eciency of biological and artifical agents in general learning environments.