Learn Fast or Multitask Well: First Steps towards a Normative Theory of Multitasking

Abstract

One of the most striking limitations of human cognition is our ability to execute some tasks simultaneously but not others. Recent work has identified overlap between task processing pathways in neural architectures as a limiting factor for multitasking performance in neural systems. This work also suggests that the brain might face a fundamental computational trade-off between learning efficiency that is gained through the use of shared representations and multitasking performance that is achieved separating the representations on which the tasks rely. According to this view, the brain faces an economic decision problem between the use of shared representations for faster learning and the use of separate representations for better multitasking performance.

Here we analyze the solution to this problem by describing the behaviour of an ideal Bayesian agent seeking to maximize their expected reward by learning either shared or separate task repre- sentations. We investigate the agent’s behaviour under different parameter settings and show that over a large parameter space the agent prefers to sacrifice long-run optimality (through higher mul- titasking performance) in favour of short-term reward (through a faster learning rate). Furthermore, we construct a general mathematical framework in which questions about the optimal behaviour of such rational agents can be analytically phrased for a wide variety of different environments. This allows us to formalize our intuitions about the nature of the learning efficiency-multitasking performance trade-off and to explore subsequent lines of inquiry in a precise fashion.