Neuro-Symbolic Artificial Intelligence: Improving upon Neural Algorithms Using Structured Knowledge

Despite impressive recent advances in deep learning, differences between the underlying mechanisms of gradient-based methods and human intelligence prevent the safe, effective application of neural networks to many tasks. We observe that neural networks are not robust, fail to generalize to examples outside the training distribution, cannot be easily interpreted by human experts, and perform poorly on combinatorial tasks. In particular, neural models are highly sensitive to small modifications to the target task or injection of malicious perturbations. Moreover, such models are very data- and compute-intensive, requiring far more training points and time to learn many tasks than humans need to achieve similar performance.

This research proposes the use of neuro-symbolic methods, which leverage task-specific symbolic information, to combat the shortcomings of neural-only methods on several challenging problems, including 1) automated verification of hardware systems, 2) adversarial robustness of computer vision models, and 3) transfer reinforcement learning.