Researchers Combine Neural Networks and Symbolic Logic for AI Common Sense
Artificial intelligence struggles with basic common-sense reasoning, as shown when advanced language models fail simple physical world questions. A new approach merging deep learning with symbolic logic offers promising progress toward solving this decades-old challenge.
Artificial intelligence systems face a major hurdle in common-sense reasoning, a problem that puzzles researchers for over fifty years. Advanced language models like GPT-2 generate impressive prose but fail spectacularly at basic physical world questions, such as correctly predicting that dropping a match on logs starts a fire instead of an "irc channel."
This persistent failure highlights the limitations of purely statistical deep learning approaches that mimic language without truly understanding the world. While these neural networks excel at pattern recognition and text generation, they lack the foundational knowledge humans use to make mundane inferences about cause and effect.
To overcome this barrier, researchers now combine neural networks with symbolic logic, merging two historically disparate lines of AI thinking. This hybrid approach integrates the pattern-matching strengths of deep learning with the rigid rule-based reasoning of symbolic systems, yielding important progress in the quest to give computers genuine common sense.