OATML

Continual learning is the process of developing new abilities while retaining existing ones. Sequential Bayesian inference over predictive functions is a natural framework for doing this, but applying it to deep neural networks is challenging in practice. To address the drawbacks of existing approaches, we formulate continual learning as sequential function-space variational inference. From this formulation, we derive a tractable variational objective that explicitly encourages a neural network to fit data from new tasks while also matching posterior distributions over functions inferred from previous tasks. Crucially, the objective is expressed purely in terms of predictive functions. This way, the parameters of the neural network can vary widely during training, allowing easier adaptation to new tasks than is possible with techniques that directly regularize parameters. We demonstrate that, across a range of task sequences, neural networks trained via sequential function-space va... [full abstract]