第71回統計的機械学習セミナー / The 71st Statistical Machine Learning Seminar
- 【Date & Time】
- December 12th (Friday), 2025 11:00 - 12:00
Admission Free, No Booking Necessary
- 【Place】
- Seminar Room 5 (3rd floor, D313/314), The Institute of Statistical Mathematics
- 【Speaker】
- Guillaume Braun (RIKEN AIP)
- 【Title】
- Neural Network Dynamics on XOR-Type Models
- 【Abstract】
-
The XOR function is one of the earliest examples of a nonlinear classification task that cannot be learned by a perceptron, and for which the introduction of a larger neural network becomes essential. Due to its structural simplicity and symmetry, XOR has become a useful test bed for analyzing neural networks in regimes where feature learning is required and classical linear methods are ineffective. In recent years, several variants of XOR-type models have been introduced to study different aspects of neural network dynamics: benign overfitting in XOR-based Gaussian mixture models, simultaneous learning of interacting features in overparameterized networks (e.g., the Glasgow two-phase dynamics), and classification in settings with no margin using Gaussian XOR. The goal of this presentation is to provide a concise overview of these developments, explain the mechanisms they uncover, and discuss potential extensions of XOR-type models.
