Nowadays, using known algorithms to solve a problem or leverging advances in deep learning to tackle it are seen as two orthogonal manners to reach a solution. However, many problems could benefit instead from a combination of the two.
Here, I will present the Neural Algorithmic Reasoning (NAR) direction and how it aims to combine the guarantees and generalization power of the classical algorithms with the adaptability and possibility of working directly with raw data of neural networks. Moreover, I will introduce NAR's successful application to the field of reinforcement learning.
More precisely, I will summarise how a neural network can imitate a dynamic programming planning algorithm, that can guarantee finding optimal solutions, resulting in performance gains especially in low-data environments.
If you’re looking at solving a real-world problem and you think there are heuristics that could be helpful, but not fully solving it and/or that require the data to be thoroughly denoised before being useful, I would say this talk is for you!
About the speaker:
PhD student in Machine Learning at Mila with Prof Jian Tang. I am broadly interested in how learning can be improved through the use of graph representations, having previously worked on algorithmic alignment for implicit planning and applications to biotechnology (drug discovery and drug combinations).
23 апр 2024