Foundations of Unsupervised Neural Induction: Learning and Reasoning from Limited Data

Abstract

Modern machine learning systems achieve remarkable performance at scale, yet they struggle to learn and generalize from small amounts of structured data. This limitation is especially evident in domains such as scientific discovery, education, and design, where data are scarce, noisy, and governed by rich compositional structure.In this talk, I present a research program aimed at enabling neural systems to acquire and manipulate structured knowledge directly from data. I focus on three directions addressing key obstacles. First, I show that modern self-supervised models can exhibit structure collapse, a failure mode in which representations lose essential structural diversity, and introduce a framework that preserves compositional structure during training. Second, I study how to integrate symbolic knowledge into neural architectures, presenting a method based on knowledge compilation that yields a new class of generative models with controllable generation, improved data efficiency, and generalization to new tasks without retraining. Third, I introduce an abstraction language for Boolean structures that enables compact representations and efficient manipulation, supporting scalable reasoning on modern parallel hardware.In the second part of the talk, I outline a broader research agenda toward unsupervised neural induction, a paradigm in which learning, abstraction, and reasoning emerge jointly from raw data. This perspective aims to bridge machine learning, symbolic reasoning, and cognitive science, enabling more reliable, interpretable, and data-efficient AI systems.

Bio

Emanuele Sansone is a Postdoctoral Fellow at MIT (CSAIL) and KU Leuven (ESAT). His research lies at the intersection of unsupervised learning and mathematical logic, with the goal of enabling machines to acquire and reason over structured knowledge. He is the recipient of a Marie Skłodowska-Curie Global Fellowship and leads an international effort to develop a benchmark for origami synthesis.