Neurosymbolic AI

Explore the cutting-edge field of neurosymbolic AI in this lecture from MIT's Introduction to Deep Learning course. Delve into the evolution of artificial intelligence, examining why current AI systems are considered "narrow" and the challenges they face with out-of-distribution performance and adversarial examples. Learn about the differences between neural networks and symbolic AI, and discover how combining these approaches in neurosymbolic AI can lead to more robust and generalizable systems. Gain insights into the MIT-IBM Watson AI Lab's research and understand the potential advantages of integrating symbolic reasoning with deep learning. Conclude with an overview of advanced concepts like CLEVERER and a summary of key takeaways in this comprehensive exploration of hybrid AI approaches.

- Introduction
- Evolution of AI
- MIT-IBM Watson AI Lab
- Why is AI today "narrow"?
- Out-of-distribution performance
- ObjectNet
- Adversarial examples
- When does deep learning struggle?
- Neural networks vs symbolic AI
- Neurosymbolic AI
- Advantages of combining symbolic AI
- CLEVERER and more
- Summary