INGI Seminar - Expressive classification rule learning with an emphasis on learning from sequential data

Machine learning, specifically neural networks, have seen significant advancements in classification tasks across various fields. Although they achieve high accuracy, their decision-making process is often incomprehensible to humans.
Rule-based methods, on the other hand, are interpretable and widely used in industries with Business Rules Management Systems.
In practice, however, these rules are written manually by experts and cannot easily be replaced by learned rules because of their limited complexity when obtained with existing methods.
In this work, we propose an end-to-end neural-based approach to learn expressive rules for classification problems. Different levels of expressiveness in rules are presented, implemented, and evaluated.
First, the learning of basic disjunctive normal form (AND-OR) with a neural network is studied.
Second, solutions to support sequential data are introduced with a recursive and a convolutional approach.
Finally, the model is extended to learn more expressive rules with predefined aggregation functions and overall complex rules.

Marine Collery is currently a Research Engineer at IBM France Lab. She recently defended her PhD on expressive classification rule learning with a neuro-symbolic approach at Institut Polytechnique de Paris (also held at Inria and IBM France Lab). Before doing her PhD, Marine worked as a Software Engineer at IBM on business automation products where her background in engineering (MSc - Grenoble INP Ensimag) and in machine learning (double degree MSc - KTH) contributed to the development of innovative solutions and features.

The seminar can also be followed on Teams

Sandwiches will be provided, please fill in the form. Deadline Day D 9:00 am