The best-laid models of mice and men: Towards a holistic characterisation of animal behaviour

Abstract

The mathematical modelling of behaviour enables the formalisation of theories of cognition. Removed from the details of neural implementation, we can abstractly reason about properties of the algorithms employed by the brain which transform the presented inputs into the observed actions. However, there are a number of complications, both in behaviour itself and the process of modelling it, which impede a comprehensive characterisation of behaviour. In this thesis, we present two separate modelling approaches which deal with some of these issues. We showcase these frameworks on the International Brain Laboratory (IBL) data of over 100 mice performing a perceptual decision- making task. Mice learn the basic contingencies of this task over a number of sessions and many thousands of trials. Afterwards, the task gains a biased block structure, requiring the animals to track this hidden state to improve their task performance. We first build a highly flexible model which deals with the issues of non- stationary behaviour due to learning and motivation, along with individual differences. This is achieved using an infinite hidden Markov model (iHMM) which provides a state based description of behaviour, with a non-parametric Bayesian structure. The latter allows for the introduction of new states in response to drastic changes in behaviour (such as learning through a sud- den insight or motivational fluctuations). We fit this model to individuals independently, exploiting automated complexity control. Dynamics in the char- acterisation of the behavioural states additionally imbue the model with the capacity to capture gradual learning. This allows us to identify distinct stages of learning which are present throughout the population of IBL mice. We also find substantial inter-individual differences in our model-based characterisations, and quantify the limited predictability of the course of learning. The second model we present uses neural networks to overcome the inherent rigidities of models such as the iHMM, by progressively removing restrictions from the class of modellable functions. This amounts to hybridising the neural networks with a classical model of expert mouse behaviour on our task, to maintain interpretability. We use this to find a simple extension of the classical model which outperforms it, and thus provides a powerful but interpretable full model of task behaviour. Amongst other insights, it shows how motivational fluctuations represent a substantial source of behavioural variability for which any complete model will have to account. We thus provide tools which bring the field closer to a holistic modelling approach of animal behaviour.