Intelligence from Learnable Novelty (arxiv.org)

Researchers have introduced a groundbreaking concept termed "learnable novelty," which integrates various aspects of intelligence across fields like machine learning, statistics, and adaptive systems. This approach addresses the dichotomy between novelty search—focused on discovering surprises—and the free-energy principle, which typically seeks to minimize them. By establishing a closed-form estimator of learnable novelty using a simple, differentiable reservoir computer, the authors claim to unify previously fragmented objectives like complexity generation, abstraction, and exploration in AI. This framework is significant for the AI/ML community as it allows for unsupervised learning that recovers complexity classifications and boosts the performance of neural cellular automata. Notably, it enhances image representation for digit classification on the MNIST dataset without labels and provides intrinsic rewards for reinforcement learning agents, improving their exploration strategies in a wide array of environments. This innovative approach offers a singular, quantifiable way to understand intelligence, promising to reshape how researchers approach the development of autonomous systems and complex algorithms.