Opening Inference Methods for Deep Learning
Inference Methods for Deep Learning
- Length
- 1 h 23 min
- Level
- Graduate
- Recorded
- MIT2026
Chapter 1 of 10: Before the lecture
Isola traces test-time search from Deep Blue and AlphaGo to o1, then weighs best-of-N, beam search and verifiers.
Inference in deep learning means more than one forward pass: it covers test-time compute, search and reasoning strategies. Isola, in session 23 of MIT 6.7960, first contrasts the statistical sense of inference with the machine-learning one. He traces look-ahead search from Deep Blue and AlphaGo to the test-time compute plots of OpenAI's o1. The core material sets teacher forcing against greedy sampling, shows why greedy decoding misses the most probable sequence, and covers best-of-N, beam search, chain-of-thought prompting and verification. Later cases include steering VQGAN with CLIP, in-context learning as gradient descent in linear settings, and test-time training.
Inference in deep learning means more than one forward pass: it covers test-time compute, search and reasoning strategies. Isola, in session 23 of MIT 6.7960, first contrasts the statistical sense of inference with the machine-learning one. He traces look-ahead search from Deep Blue and AlphaGo to the test-time compute plots of OpenAI's o1. The core material sets teacher forcing against greedy sampling, shows why greedy decoding misses the most probable sequence, and covers best-of-N, beam search, chain-of-thought prompting and verification. Later cases include steering VQGAN with CLIP, in-context learning as gradient descent in linear settings, and test-time training.
The Bitter Lesson
AlphaGo
Scaling Laws of Board Games
This lecture
Lecture 23 of 24 of Deep Learning.
Choose a node to see how it connects to this lecture. What comes before is on the leftat the top. What comes after is on the rightat the bottom. The ring is what it teaches. People and places are above. Other lectures and books are below.
Press a chosen node again to open it.
Your comment is kept while you sign up.Sign in
Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
Isola et al. 2011, Khosla et al. 2015
GAnalyse: Toward Visual Definitions of Cognitive Image Properties
Zhao et al. '21
Transformers learn in-context by gradient descent
Test-Time Training
The Surprising Effectiveness of Test-Time Training for Abstract Reasoning
Mastering the game of Go with deep neural networks and tree search
STaR: Self-Taught Reasoner Bootstrapping Reasoning With Reasoning
Nature journal
Generating data that optimizes..., Forgettable and memorable
The Surprising Effectiveness...
STaR