Gary Marcus: Toward a Hybrid of Deep Learning and Symbolic AI | Lex Fridman Podcast #43

Symbolic AI

Gary Marcus emphasizes the enduring relevance of symbolic AI in modern computing, arguing that it complements deep learning by handling tasks unsuitable for machine learning alone. He illustrates this with examples like phone operating systems, which require precise symbol manipulation rather than machine learning-based approaches 1. Marcus advocates for a hybrid intelligence approach, combining the strengths of both symbolic AI and deep learning to create more robust systems 2.

We need new technologies that are going to draw some of the strengths of both the expert systems and the deep learning, but are going to find new ways to synthesize them.

--- Gary Marcus

This hybrid model could leverage deep learning for perceptual tasks while using symbolic AI for logical reasoning and inference.

   

AI Limitations

Marcus critiques the current machine learning paradigm, highlighting its limitations in abstract reasoning and common sense understanding. He points out that while deep learning excels in perceptual classification, it struggles with representing abstract knowledge, such as the concept of containers or causality 3. Marcus suggests revisiting older AI concepts, like symbolic manipulation, which could benefit from modern computational power and data availability 4.

It could be that symbol manipulation per se with modern amounts of data and compute might be great.

--- Gary Marcus

By integrating these older ideas with new technologies, AI could achieve a more comprehensive understanding of the world.