UCLA

Despite the recent remarkable advances in deep learning, we are still far from building machines with human-like general intelligence, for instance, understanding the world in a fast, structured, and generalizable way. The dominant stream in contemporary AI hopes to achieve human-level performance via purely data-driven methods, \ie, fitting deep neural networks on a massive amount of training data. However, these methods are often trapped in a dilemma of ``big data, small tasks'', and are hard to interpret and generalize.

In this dissertation, we seek a unified framework for general intelligence by integrating connectionism and symbolism in a neuro-symbolic system. We argue that (i) \textbf{Neural Network} is excellent at imitating human perception from raw signals, (ii) \textbf{Grammar} provides a universal approach to construct a holistic structured representation of the world, and (iii) \textbf{Symbolic Reasoning} forms a principled basis to incorporate commonsense knowledge and perform complex reasoning. Therefore, we propose a neural-symbolic framework by using grammar as the bridge to connect neural networks and symbolic reasoning. The learning of such a neural-symbolic framework mimics human’s ability to learn from failures via abductive reasoning and requires very little supervision. We have developed benchmarks, algorithms, and practices, across vision and language, from synthetic environments to real-world scenarios, to realize such a unified framework. We hope such a unified framework can contribute to the long-term goal of building general artificial intelligence like humans.