Abstraction Learning
Summary
Abstraction learning is proposed as a key approach to bridging the gap between artificial and human intelligence. Unlike previous AI research that either relied on human-specified abstractions or used abstraction as a qualitative explanation for model behavior, this approach aims to directly learn abstractions. The authors identify three main challenges in abstraction learning: representation, objective function, and learning algorithm. To address these challenges, they introduce ONE (Optimization via Network Evolution), a framework that utilizes a partition structure with pre-allocated abstraction neurons, formulates abstraction learning as a constrained optimization problem incorporating abstraction properties, and employs a network evolution algorithm for learning. Experiments on the MNIST dataset demonstrate that ONE exhibits elements of human-like intelligence, including low energy consumption, knowledge sharing, and lifelong learning capabilities.