Learning Explanatory Rules from Noisy Data - Richard Evans, DeepMind

Explore a 26-minute conference talk on learning explanatory rules from noisy data, presented by Richard Evans from DeepMind at the Alan Turing Institute. Delve into the challenges of artificial neural networks and their tendency to overfit, contrasting them with logic programming methods like Inductive Logic Programming (ILP). Discover the proposed Differentiable Inductive Logic framework (∂ILP), which combines the strengths of both approaches to handle noisy data and non-symbolic domains. Examine various topics including learning procedures from examples, symbolic program synthesis, neural program induction and synthesis, and the application of these concepts to tasks such as graph cyclicity and Fizz-Buzz. Gain insights into handling mislabelled data, learning rules from ambiguous data, and comparing image and symbolic generalisation with MLP baselines.