Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas 336 | Anil Ananthaswamy on the Mathematics of Neural Nets and AI
121 snips
Nov 24, 2025 
Anil Ananthaswamy, a science writer and former New Scientist editor, explores the intricate mathematics behind artificial intelligence. He dives into topics like neural networks, the challenges of training models on planetary motion, and the significance of algorithms like backpropagation. The conversation highlights the elegance of the perceptron proof and discusses the complexities of high-dimensional spaces and transformer architectures. Anil also reflects on the implications of AI's mathematical foundations for scientific discovery.
AI Snips
Chapters
Books
Transcript
Episode notes
Self-Taught Deep Learning Project
- Anil Ananthaswamy taught himself deep learning during an MIT Knight Fellowship and built neural nets to predict planetary positions from simulated data.
- He realized networks need far more data than human scientists like Kepler, motivating his deeper study of the math behind ML.
Perceptron Guarantees Linear Separation
- The perceptron is a linear classifier that separates data via a hyperplane if data are linearly separable.
- The perceptron convergence proof guarantees finding such a plane in finite time when separability holds.
Weekend-Built Hardware Neuron
- Bernie Widrow and Ted Hoff built an early hardware artificial neuron using adaptive filter ideas and even assembled circuitry over a weekend.
- Their least mean squares algorithm anticipated stochastic gradient descent and influenced backpropagation's lineage.
