Machine Learning Street Talk (MLST) #51 Francois Chollet - Intelligence and Generalisation
10 snips
Apr 16, 2021 Francois Chollet, the genius behind Keras and author of 'Deep Learning with Python,' shares his profound insights on intelligence as generalization. He challenges the limitations of neural networks, arguing they struggle with reasoning and planning. The discussion explores the future of AI, emphasizing the need for program synthesis and the integration of discrete methods. Chollet dives into the nuances of generalization and abstraction, highlighting how these concepts can shape a new era in AI innovation. Expect a fascinating journey through the complexities of intelligence!
AI Snips
Chapters
Books
Transcript
Episode notes
Deep Learning as Interpolation
- Deep learning models generalize through interpolation on high-dimensional manifolds.
- They function like low-quality sensitive hash tables with a notion of distance between data points.
LSTM for Multiplication
- An LSTM was trained to multiply three-digit numbers, learning the algorithm itself.
- However, it exhibited glitches, required extensive data, and lacked broad generalization.
Hybrid Problem Solving
- Many problems blend continuous and discrete structures, requiring hybrid solutions.
- Program search, potentially guided by deep learning, offers a promising approach.
