Introduction

When learning from structured data, special attention has to be paid to the models used. Indeed, designing machine learning models requires to think of the invariants to be learned [Arjovsky et al., 2019], and either encode them in the model or design the model so that it is able to discover such invariants and encode them.

In this course, we will focus on time series and will dig into two main ways of encoding / learning these invariants. First, we will cover the design of alignment-based metrics that tackle the problem of (temporal) localization invariance. Standard similarity measures will be introduced and their use at the core of machine learning models will be discussed. Second, we will discuss standard neural network architectures and the kind of invariants they encode.

References

ABGLP19

Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization. 2019. arXiv:1907.02893.