Cooperation of supervised and unsupervised learning in representation formation and task acquisition

A central question in both neuroscience and machine learning is how animals are able to learn complex tasks from very limited supervision. This capability is widely believed to be supported by various forms of unsupervised learning, operating at both the level of learning rules and learning objectives. However, it remains an underexplored question how these different learning mechanisms interact within the brain. This thesis investigates this question from a representation learning perspective, focusing on the relationship between biologically realistic representations and task performance. First, we examine the interaction between a Hebbian-based learning rule and a gradient-based learning rule in a visual categorisation task. We find that, although unsupervised learning produces representations that resemble those observed in the early visual cortex, it slightly hinders task performance and provides only marginal benefits for generalisation. Nevertheless, the two learning rules can operate simultaneously with minimal coordination and without significant performance loss, suggesting the potential for cooperation between different learning rules in tasks where generalisation and robustness are important. The results also indicate that gradient-induced neural plasticity associated with task optimisation may be difficult to detect experimentally. We then investigate the interaction between predictive learning and reinforcement learning in navigation tasks. We find that predictive learning can significantly enhance reinforcement learning, but the two objectives are strongly competing, and the benefits of predictive learning are limited without proper coordination. We compare different integration approaches and show that a staged learning procedure is particularly effective: the agent first explores the environment using predictive learning alone, and then jointly optimises predictive learning and reinforcement learning during task acquisition. This procedure allows for the emergence and preservation of hippocampal-like representations, which are directly associated with the highest task performance. These representations improve the stability and speed of reward acquisition, enhance information retention during sequential multi-task learning, and facilitate rapid acquisition of new tasks that share a common structure. Overall, this work demonstrates both the challenges and the potential of coordinating supervised and unsupervised learning in biological and artificial systems, and shows that representation analysis is a useful tool for understanding task behaviour and designing better coordination strategies.