Fireside Chat with Aparna Ramani & Yann LeCun
summary
Argues that self-supervised learning — “fill in the blanks” pretraining becomes the default: massive unlabeled pretraining (example: ~a billion Instagram images) followed by light supervised fine-tuning yields few-shot classifiers. Notes successes in text and protein folding but failures for raw video and images because of continuous uncertainty. Proposes replacing full probabilistic generative modeling with energy-based models and joint embedding architectures that predict in abstract space, and describes Transformers as high-capacity associative memories enabling large unified models.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 0:49Speakers and AI roles
- 1:32Yann LeCun on intelligence
- 2:58AI future and winters
- 5:26Deep learning limits and breakthroughs
- 6:10Self-supervised learning
- 9:05Why self-supervised learning matters
- 11:55AI/ML challenges in data systems
- 15:29Math breakthroughs for AI
- 18:22Q&A on data and compute
- 20:13Quantum computing's AI impact
- 21:57Why Transformers work
- 24:07Grounded AI and human-level intelligence
- 25:33AI applications in our lifetime
- 28:40AI's future summary
related talks
01h 07mDiscussion Session
35mFireside Chat
26mA Shallow Introduction to Self-Attention
32mBeyond Algorithms – on Human(oid) Intelligence
25mEmbeddings, Transformers, RLHF: Three key ideas to understand ChatGPT
26mGenerative AI: Get a Grip … and a Look Under the Hood
51mLarge Language Models for Real-World Applications - A Gentle Intro
01h 30mToward Understanding In-context Learning