Explore the intersection of statistics and computer science in machine learning through this 42-minute conference talk by Jean-Philippe Vert from Google Brain. Delve into innovative approaches for embedding permutations and relaxing ranking operators to integrate them into differentiable machine learning architectures. Discover how to analyze and predict preferences using continuous space representations of discrete combinatorial objects. Examine the SUQUAN and Kendall embeddings, optimal transport techniques, and applications such as soft top-k loss and learning to sort. Gain insights into the cross-fertilization between statistics and computer science in the era of Big Data, and understand the algorithmic paradigms underpinning modern machine learning.
Learning From Ranks, Learning to Rank - Jean-Philippe Vert, Google Brain
Alan Turing Institute via YouTube
Overview
Syllabus
Intro
Differentiable programming
Beyond images and strings
What if inputs or outputs are permutations?
Examples
Goals
Permutations as inputs
SUQUAN embedding (Le Morvan and Vert, 2017)
Supervised ON (SUQUAN)
Experiments: CIFAR-10
Limits of the SUQUAN embedding
The Kendall embedding (Jiao and Vert, 2015, 2017)
Geometry of the embedding
Kendall and Mallows kernels
Applications
Remark
Permutations as intermediate / output?
Optimal transport (OT)
Differentiable permutation matrix
Differentiable sort and rank
Soft quantization and soft quantiles
Application: soft top-k loss
Application: learning to sort
Conclusion
Taught by
Alan Turing Institute
