Modular & Causal Knowledge Representation for Lifelong Learning - Yoshua Bengio - CoLLAs 2022

Описание: Abstract:
How can what has been learned on previous tasks generalize quickly to new tasks or changes in distribution? The study of conscious processing in human brains (and the window into it given by natural language) suggests that we are able to decompose high-level verbalizable knowledge into reusable components (roughly corresponding to words and phrases). This has stimulated research in modular neural networks where attention mechanisms can be used to dynamically select which modules should be brought to bear in a given new context. Another source of inspiration for tackling this challenge is the body of research into causality, where changes in tasks and distributions are viewed as interventions. The crucial insight is that we need to learn to separate (somewhat like in meta-learning) what is stable across changes in distribution,environments or tasks and what may be separate to each of them or changing in non-stationary ways in time. From a causal perspective what is stable are the reusable causal mechanisms, along with the inference machinery to make probabilistic guesses about the appropriate combination of mechanisms (maybe seen as a graph) in a particular new context. What may change with time are the interventions and other random variables which are those that yield more directly to observations. If interventions are not observed (we do not have labels for fully explaining the changes in tasks in terms of the underlying modules and causal variables) we would ideally like to estimate the Bayesian posterior over the interventions, given whatever is observed.This research approach raises many interesting research questions ranging from Bayesian inference and identifiability to causal discovery, representation learning and out-of-distribution generalization and adaptation, which will be discussed in the presentation.