Alternating inference and learning: a thalamocortical model for continual and transfer learning

Animals thrive in a constantly changing environment and leverage the temporal structure to learn well-factorized causal representations. In contrast, traditional neural networks have no representation of these changes and suffer from forgetting. In the prefrontal-thalamic circuit, reciprocal feedback connections produce a compressed thalamic representation of the computations performed in the prefrontal cortex. In inspiration, we use a neural network with a separate latent task representation and introduce a simple algorithm that uses optimization to rapidly infer and generate representations of the current task. The algorithm alternates between updating the neural network weights and the latent task embedding, allowing the agent to parse the stream of temporal experience into discrete tasks and organize learning about them. The algorithm achieves competitive end average accuracy on a continual learning benchmark, but importantly, by requiring the model to adapt through inferring tasks (latent updates), it organizes task knowledge into structures and the thalamic task embedding becomes a cognitive interface to control them. Tasks later in the sequence can be solved through knowledge transfer as they become reachable within the well-factorized thalamic latent space. The algorithm adapts to a stream of unlabeled tasks through task inference, rather than re-learning, consistent with accounts of animal behavior, and proposes feedback control as a fundamental computation in the corticothalamic projections in the brain.