6.7960 Deep Learning (MIT, Fall 2026)
by MIT (Phillip Isola, Kaiming He)
MIT's graduate deep learning course, co-taught by Kaiming He and Phillip Isola, covering architectures, representation learning, scaling laws and generative models.
Overview
6.7960 Deep Learning is MIT EECS's graduate deep learning course. The Fall 2026 edition is co-taught by Phillip Isola and Kaiming He, lead author of ResNet, with a staff of about twenty TAs. It meets twice a week from September 10 to December 15, 2026. The public schedule has 24 lecture slots, three of them guest lectures, plus a midterm and a final. It starts with how to train a neural net and approximation theory. It then covers architectures (ConvNets, sequence modeling, transformers), generalization theory and going deep with neural networks. Next is representation learning in three lectures: methods, weight-space geometry and information theory. Foundation models get two lectures, on pre-training and on scaling laws. Generative models get three: basics, VAEs and GANs, and diffusion and flows. Later lectures cover out-of-distribution generalization, transfer learning for models and data, inference-time algorithms, evaluation, and applying deep learning to your own problems. Enrolled students complete five two-week problem sets mixing derivations and coding, which get AI-agent feedback. The midterm and final are closed-book. Slides are linked publicly from the schedule. Fall 2026 recordings are Canvas-only, so self-learners should pair the slides with the Fall 2024 edition on MIT OpenCourseWare, which has full videos, notes, problem sets and projects under CC BY-NC-SA 4.0. Readings come from the free books Foundations of Computer Vision and Understanding Deep Learning.
At a Glance
- Topic
- ML
- Level
- Advanced
- Format
- Course
- Cost
- Free
- Duration
- 14 weeks, Sep 10 – Dec 15, 2026, with 24 lecture slots; about 90 hours of coursework per csdiy.wiki's estimate
- Provider
- MIT (Phillip Isola, Kaiming He)
- Hands-on
- Yes — code/exercises
- Certificate
- None
What You’ll Learn
- ✓How to train a neural network well, and the approximation theory behind why it works
- ✓How ConvNets, sequence models and transformers differ in the inductive biases they encode
- ✓What generalization theory, including out-of-distribution generalization, says about when deep nets work
- ✓How representation learning methods work, viewed through weight-space geometry and information theory
- ✓How foundation models are pre-trained and what scaling laws predict about compute and data
- ✓How VAEs, GANs, diffusion and flow-matching generative models are built and trained
- ✓How transfer learning and inference-time algorithms get more out of a trained model
- ✓How to evaluate deep learning systems and apply them to your own problems
Highlights
- •Co-taught by Kaiming He (ResNet) and Phillip Isola, who split the lectures between them
- •The 2026 syllabus adds foundation-model pre-training, scaling laws, inference-time algorithms and evaluation to the classic theory core
- •Self-learners can pair the public 2026 slides with the full Fall 2024 OCW edition (videos, notes, psets, projects)
- •Balances vision and language instead of specializing in one, which csdiy.wiki singles out as its main strength
Who It’s For
Best For
- ✓ML engineers who use transformers daily and want the theory underneath them
- ✓Engineers moving from applied LLM work into model training or research
- ✓Graduate students or strong self-learners preparing for research on foundation or generative models
- ✓Practitioners who want a rigorous treatment of diffusion and flow models alongside LLMs
Prerequisites
- •Probability and statistics at the level of MIT 18.05
- •An introductory machine learning course (MIT 6.3720, 6.3900 or 6.C01 equivalent)
- •Linear algebra and Python programming for the coding parts of the problem sets
FAQ
What is 6.7960 Deep Learning (MIT, Fall 2026)?
6.7960 is MIT's graduate-level deep learning course, taught in Fall 2026 by Phillip Isola and Kaiming He (known for ResNet). It is for engineers with probability and ML foundations who want the theory behind modern models. It covers why architectures work, how foundation models scale, and how diffusion and flow models are built.
Is 6.7960 Deep Learning (MIT, Fall 2026) free?
6.7960 Deep Learning (MIT, Fall 2026) is free to access.
What level is 6.7960 Deep Learning (MIT, Fall 2026) for?
6.7960 Deep Learning (MIT, Fall 2026) is aimed at a advanced audience. Recommended background: Probability and statistics at the level of MIT 18.05, An introductory machine learning course (MIT 6.3720, 6.3900 or 6.C01 equivalent), Linear algebra and Python programming for the coding parts of the problem sets.
How long does 6.7960 Deep Learning (MIT, Fall 2026) take?
Expect roughly 14 weeks, Sep 10 – Dec 15, 2026, with 24 lecture slots; about 90 hours of coursework per csdiy.wiki's estimate. Most learners work through it at their own pace.
What will I learn from 6.7960 Deep Learning (MIT, Fall 2026)?
You'll learn: How to train a neural network well, and the approximation theory behind why it works; How ConvNets, sequence models and transformers differ in the inductive biases they encode; What generalization theory, including out-of-distribution generalization, says about when deep nets work; How representation learning methods work, viewed through weight-space geometry and information theory; How foundation models are pre-trained and what scaling laws predict about compute and data; How VAEs, GANs, diffusion and flow-matching generative models are built and trained; How transfer learning and inference-time algorithms get more out of a trained model; How to evaluate deep learning systems and apply them to your own problems.
Topics
Sources
This page was written from 3 sources, 2 on domains other than deeplearning6-7960.github.io.