Generalization on the unseen domain in real case


Event details

Date 16.02.2024
Hour 10:0012:00
Speaker Denys Pushkin
Category Conferences - Seminars
EDIC candidacy exam
Exam president: Prof. Lénaïc Chizat
Thesis advisor: Prof. Emmanuel Abbé
Co-examiner: Prof. Nicolas Flammarion

This write-up summarizes three papers in learning
theory field. The first studies the generalization on the unseen
domain in Boolean domain and shows that a number of model
architechtures follow the same bias on the unseed. The second
introduces a dataset of math problems, which allows to test
models generalization ability across different axes of complexity.
The last studies the connection of random features and nueral
tangent kernel models with 2-layers neural network.

Background papers
1. Emmanuel Abbe, Samy Bengio, Aryo Lotfi, Kevin Rizk "Generalization on the Unseen, Logic Reasoning and Degree Curriculum"
2. Behrooz Ghorbani, Song Mei, Theodor Misiakiewicz, Andrea Montanari "Linearized two-layers neural networks in high dimension"
3. David Saxton, Edward Grefenstette, Felix Hill, Pushmeet Kohli "ANALYSING MATHEMATICAL REASONING ABILITIES OF NEURAL MODELS"

Practical information

  • General public
  • Free


EDIC candidacy exam