Network dynamics implement optimal inference in a flexible timing task

The brain's ability to produce reliable behaviour in a dynamic world is critical to survival. Humans and animals can quickly adapt their behaviour to subtle environmental changes, suggesting reliance on neuronal activity rather than long term synaptic plasticity. However, it is unclear how a neural network can quickly change its response based on subtle changes in its inputs. We developed a novel GO/NO-GO evidence accumulation task that requires fast responses to hidden-state changes under varying uncertainty levels. The goal is to detect transitions between `safe' and `unsafe' states. In the safe state, a constant GO cue is presented. In the unsafe state, a NO-GO cue is presented with intermittent random misleading GO cues. The probability of a misleading GO cue in the unsafe state is a parameter we control, which defined the noise level in the task. The agent must strike a balance between speed and accuracy: acting prematurely may result in an erroneous trial while waiting to remove uncertainty delays the reward---both reduce the overall reward rate. Importantly, the optimal waiting time depends on the noise level, which we change throughout the experiment. The challenge is that the agent needs to estimate the noise level to evaluate the hidden state and vice versa.

We trained mice and artificial recurrent neural networks (ANNs); both quickly adjusted their waiting times to the noise level. We analyzed the trained ANNs' activity and found that networks represent the hidden-state probability and noise level in orthogonal directions. Furthermore, the noise estimation matches that of an ideal Bayesian observer, suggesting that the network learned the structure of the task. Our results show that neural networks can learn flexible task structures and adapt their response online. This study advances our understanding of how neural networks may represent complex and flexible cognitive tasks.