2025 workshop program
Exploring Compositional Generalization of Neural Networks through Synthetic Experiments
The work
Abstract
Interpreting neural networks remains a significant challenge due to their complexity. Approaches range from fine-grained mechanistic analyses to evaluating broad performance benchmarks. This tutorial introduces an intermediate strategy: understanding neural network behavior via controlled synthetic experiments. Using a synthetic toy model of spectral detection—where spectral abundances map directly to generated spectra—we will train neural networks to infer abundances from spectra alone. By systematically varying training data distributions, we will analyze when neural networks compositionally generalize to scenarios unseen during training. We will replicate and extend prior observations from generative diffusion models, mainly investigating generalization across multiple dimensions. Participants are encouraged to explore and test various hypotheses within this controlled experimental framework in an open ended way.
Requirements: Installed jupyter notebook or colab with numpy, matplotlib, pytorch.