🌏 中文版
Video status: Official entry or recording index only. Source details
MIT 6.7960: Deep Learning is MIT EECS's graduate-level deep learning course. In Fall 2025 it is co-taught by Sara Beery, Kaiming He, and Omar Khattab. Every one of the semester's 21 lecture decks is directly available as a Dropbox PDF, weekly readings are listed on the schedule, and even the final project's guidelines PDF is public.
This is not, however, a fully open course. Problem sets are released through Gradescope, their solutions sit on Canvas, and so do the lecture recordings — all three require MIT credentials. Under this site's four-tier rubric from the Global AI and CS Course Map, the course rates A2: partially open materials. A materials-level deep dive is possible; the gaps must be stated plainly.
This post answers four questions: what the course teaches, how the three instructors divide it, how the 21 lectures connect, and how far an outside learner can get using two free textbooks. My verification covered the complete Fall 2025 schedule, grading section, and policies on the official site; I did not open every slide deck to audit its inner pages, so descriptions of individual lectures come from the official site's own summaries.
First, decide if this is your course
The official description promises "fundamentals of deep learning, including both theory and applications," covering neural net architectures (MLPs, CNNs, RNNs, graph nets, transformers), backpropagation and automatic differentiation, learning theory and generalization in high dimensions, plus applications to computer vision, NLP, and robotics. The prerequisites are stiff: 18.05 (probability and statistics) plus one of 6.3720, 6.3900, or 6.C01 — meaning you should already have taken a machine learning or algorithms subject before walking in. It carries 3-0-9 units, and due to heavy enrollment Fall 2025 accepted no cross-registrations at all.
Compared with MIT 6.S191, which this site has also reviewed, the positioning is entirely different. 6.S191 is a nine-week bootcamp that prioritizes breadth; 6.7960 is a full-semester backbone course with proofs and theory. Generalization theory, approximation theory, and information theory — the topics bootcamps skip — each get a full lecture here. If you want to understand why things work rather than merely call APIs, this skeleton is worth walking through.
The grading structure reveals the emphasis: five psets worth 10% each (half the grade), a midterm worth 25%, and a final research project worth 25%. The exam is two hours, closed book, with a single page of handwritten notes allowed — the course expects you to internalize derivations, not look them up.
Three instructors, exactly seven lectures each
The 21 lectures split evenly among the three instructors, and the division is not turn-taking — each owns a thematic line:
| Instructor | Line | Lectures |
|---|---|---|
| Sara Beery | Engineering foundations and deployment | 1–2, 4, 6, 17–18, 20 |
| Kaiming He | Representation learning and generative models | 5, 8, 10, 13–14, 16, 21 |
| Omar Khattab | Theory and foundation models | 3, 7, 9, 11–12, 15, 19 |
Beery owns "how you train models and ship them into the real world": the opening training fundamentals, CNNs, Transformers, OOD generalization, transfer learning, and a closing lecture on evaluation. He handles "what happens inside the model": sequence modeling, three flavors of representation learning, four lectures of generative modeling, and the finale on applying deep learning to your own problems. Khattab carries theory and the LLM side: approximation theory, generalization theory, similarity-based retrieval (his own neural information retrieval specialty), plus pre-training, scaling laws, post-training as three foundation-model lectures, and inference-time algorithms.
The lineup itself tells a story. Kaiming He is the author of ResNet; Omar Khattab created ColBERT and DSPy. In Fall 2024 the course was taught by Phillip Isola, Jeremy Bernstein and Sara Beery (per the OCW video descriptions: Beery gave Lec 01, 02, 04, 10, 12, 17 and 18; Isola gave Lec 05, 06, 08, 09, 11, 14, 15 and 16; Bernstein gave Lec 03, 07 and 13); after the 2025 handover the syllabus visibly shifted toward LLMs and retrieval — the foundation-model block and inference-time algorithms are new.
A topic map of the 21 lectures
Laid out, the schedule falls into six segments:
| Segment | Lectures | Content |
|---|---|---|
| Foundations | 1–3 | DNN building blocks, SGD/backprop/autodiff, universal approximation through Barron's theorem |
| Architectures | 4–6 | CNNs (grids), RNNs/LSTMs (memory), Transformers (tokens, attention, positional codes) |
| Generalization and representation learning | 7–10 | Generalization theory, reconstruction-based representation learning, similarity-based (contrastive, InfoNCE), information theory |
| Foundation models | 11–12, 15 | Pre-training, scaling laws, post-training |
| Generative models | 13–14, 16 | Generative basics, VAE and GAN, diffusion and flows |
| Practice and deployment | 17–21 | OOD generalization, transfer learning, inference-time algorithms, evaluation, practical advice |
Three things stand out. First, architectures take only three lectures: MLPs, CNNs, RNNs, and Transformers are taught as variations on the same set of ideas, not as a tour of popular models. Second, "generalization" appears twice — Lecture 7 covers the theory, Lecture 17 covers distribution shift and robustness; the theoretical and engineering faces are kept separate. Third, LLMs are not one standalone lecture; they are decomposed into pre-training, scaling laws, post-training, and inference-time algorithms, threaded through the deep learning spine.
Three guest lectures are interleaved: Rose E Wang (OpenAI), Zongyi Li (MIT/NYU), and Jiajun Wu (Stanford), with no public slides. There is also a PyTorch tutorial whose Colab notebook is public and runnable from anywhere.
Two free textbooks are your recording substitute
Without recordings, the readings become an outside learner's main channel. The good news: most of them point to free textbooks.
- Foundations of Computer Vision by Torralba, Isola, and Freeman (MIT Press, fully readable online for free). Most required readings come from here: Lecture 1 pairs with the neural networks chapter; Lecture 2 pairs with gradient-based learning and backpropagation; CNNs and Transformers each have matching chapters.
- Understanding Deep Learning by Simon Prince (free PDF). The course site calls it "probably the best textbook devoted entirely to deep learning." UDL is not assigned chapter-by-chapter, but its chapters on inference, diffusion, and transformers cover exactly what the vision book treats lightly.
- To refresh ML basics, the official site also points to the open 6.390 notes.
The optional readings are curated seriously too. Generalization theory lists Understanding deep learning requires rethinking generalization; pre-training lists the GPT-3 paper Language Models are Few-Shot Learners alongside OLMo 2; scaling laws pairs Kaplan et al. with Chinchilla plus both sides of the "are emergent abilities a mirage?" debate. All public links — effectively a curated must-read list for modern deep learning.
The self-study method is direct: treat the official schedule as your syllabus, read the required readings before opening each deck, and when a derivation in the slides doesn't land, go back to the matching textbook chapter and redo it. The UDL author also ships a full set of notebooks, so any formula you can't parse can be verified in code.
The assignment gap, stated plainly
This is the line between A2 and A3:
| Item | Status | Notes |
|---|---|---|
| 21 lecture decks | Public | Direct Dropbox PDF links |
| Weekly readings | Public | Free textbook chapters, arXiv papers |
| Project guidelines | Public | Including grading direction and team rules |
| PyTorch tutorial | Public | Runnable Colab notebook |
| Pset questions | Not public | Released via Gradescope, requires enrollment access |
| Pset solutions | Not public | Links resolve to Canvas files |
| Lecture recordings | Not public | Canvas external tool, enrolled students only |
| TAs, office hours, feedback | Not public | In-class environment |
The final project is the one assessment an outsider can complete in full. The guidelines PDF asks for novel experimentation and visualization, teams of up to three, and explicitly provides Google Colab Pro while telling students not to plan for large compute costs — "be creative." That design is friendly to self-learners: one small, clean research question beats hoarding GPU hours.
Worth noting is the AI assistants policy: rules for ChatGPT and friends are identical to those for human assistants — asking about concepts is welcome, doing the work for you is not, and any use must be disclosed atop the pset. The course treats understanding what AI can and cannot do as part of its content, and the policy reads more honestly than most.
The late policy is instructive too: homework is accepted up to 7 days late with score multiplied by (1−n/14), and every student gets ten penalty days waived automatically. The design assumes life happens, so you don't beg case by case.
An outside self-study route
Once the gaps are acknowledged, the path forward is clear. My advice is to replace "finished reading" with "built something":
- Test the water first: open the PyTorch tutorial Colab and complete the first cell block without peeking at anything. If you're stuck, go back to the 6.390 notes before pushing on.
- Textbook first, slides second: for the first three weeks, follow the readings into visionbook's learning part (gradient-based learning, backpropagation, neural networks) alongside Lectures 1–3. After each chapter, redo the key derivation on blank paper.
- Invent small problems where psets are missing: for example, after backpropagation, implement finite-difference gradient checking for a one-layer network in NumPy. It won't impersonate a real pset, but it verifies you actually understand.
- Read the foundation-model block (Lectures 11, 12, 15, 19) with the papers: this is the newest part of the 2025 syllabus, and the optional readings — GPT-3, OLMo 2, Kaplan, Chinchilla — are the core material.
- Finish with a mini project: following the spirit of the guidelines, pick something runnable on free Colab and write it up like a Distill-style analysis. This is the one assessment you can reproduce end to end.
The formal course runs 15 weeks; budgeting 12 weeks for self-study is realistic, because without office hours, stuck-time comes out of your own schedule.
Which edition to use: 2025 as syllabus, 2024 OCW for recordings
The Fall 2024 edition, taught by Isola, Beery, and Bernstein, also published every deck — and graded 65% psets / 35% project with no midterm. Adding an exam in 2025 signals a raised bar on internalizing theory. Fall 2023, then numbered 6.s898, is an earlier ancestor.
The key difference is video. Fall 2025 recordings are locked behind Canvas, but the Fall 2024 course shipped on MIT OpenCourseWare — videos, notes, five assignments, and the project, all public. So my recommendation: use the Fall 2025 schedule and decks as your primary syllabus (freshest instructor lineup and LLM content), and watch corresponding-topic recordings from Fall 2024 on OCW where you want a lecture voice. The two years don't align perfectly — graph nets had a dedicated lecture in 2024 that left the main 2025 table — but reading them side by side shows how the same course evolves.
For background, pair this with the MIT AI/ML Course Map and the CMU 11-785 overview: the former locates 6.7960 within MIT's ladder, the latter reviews another course with open teaching and closed assignments.
OCW Fall 2024, Lecture by Lecture
OCW's Readings page lists 25 sessions. For 23 of them a recording and slides are public, and this site has a guide to each:
| Lec | OCW topic | Guide on this site |
|---|---|---|
| 1 | Introduction to Deep Learning | Lecture 1 |
| 2 | How to Train a Neural Net | Lecture 2 |
| 3 | Approximation Theory | Lecture 3 |
| 4 | Architectures: Grids | Lecture 4 |
| 5 | Architectures: Graphs | Lecture 5 |
| 6 | Generalization Theory | Lecture 6 |
| 7 | Scaling Rules for Optimization | Lecture 7 |
| 8 | Architectures: Transformers | Lecture 8 |
| 9 | Hacker's Guide to Deep Learning | Lecture 9 |
| 10 | Architectures: Memory | Lecture 10 |
| 11 | Representation Learning: Reconstruction-Based | Lecture 11 |
| 12 | Representation Learning: Similarity-Based | Lecture 12 |
| 13 | Representation Learning: Theory | Lecture 13 |
| 14 | Generative Models: Basics | Lecture 14 |
| 15 | Generative Models: Representation Learning Meets Generative Modeling | Lecture 15 |
| 16 | Generative Models: Conditional Models | Lecture 16 |
| 17 | Generalization: Out-of-Distribution (OOD) | Lecture 17 |
| 18 | Transfer Learning: Models | Lecture 18 |
| 19 | Transfer Learning: Data | Lecture 19 |
| 20 | Scaling Laws | Lecture 20; public recording covers only the first 38 minutes; the rest follows the slides |
| 21 | Language Models | Lecture 21 |
| 22 | AI for Musical Creativity | No guide: no recording or slides were published |
| 23 | Metrized Deep Learning | Lecture 23 |
| 24 | Inference Methods for Deep Learning | Lecture 24 |
| 25 | Efficient Policy Optimization Techniques for LLMs | No guide: no recording or slides were published |
Why Lectures 22 and 25 have no guide. OCW's Lecture Notes page says "Lectures 22 and 25 are not available," and the video gallery omits both; with no material to read, there is nothing to write from. OCW gives no reason. Two things can be confirmed: according to the instructors' course site, Lecture 22 (2024-11-21) was a guest lecture by Anna Huang and Lecture 25 (2024-12-10) was given by Kianté Brantley, neither of them an instructor of the course; and the course site has no slide link for these two sessions, while every other session has one. The most plausible reading is that the speakers did not release their material, but that is this site's inference, not an official statement. Being a guest lecture is not the reason in itself: Lecture 21 was also a guest lecture (Jacob Andreas), and its recording and slides are public.
The Lecture 20 recording covers only the first half. The published video is 38 minutes long (the others run about 68 to 83 minutes), and the official transcript also ends mid-sentence. The second half of that guide follows the slides only, as marked in the text.
Besides the lecture guides, the series has three topic guides that do not match a single lecture (Optimization Overview, Regularization in Practice, Modern CNN Architectures) and one Assignment 1 walkthrough.
Bottom line
6.7960 Fall 2025 is a rigorously designed A2 course: 21 public decks, curated readings, and an open project guidelines document make materials-level self-study entirely feasible, while closed psets, solutions, and current-year recordings mark the real boundary for outsiders. Substitute textbooks for lectures, fill recordings from OCW 2024, and close with a mini project — you won't get MIT credit, but you will reach the core of the course.
Course video sources
This article covers multiple lectures; choose recordings by topic and lecture from the official index. Rechecked against the live MIT OCW Fall 2024 gallery on 2026-10-10: it lists public recordings for Lec 01–21, Lec 23, Lec 24 and a PyTorch Tutorial; OCW marks Lec 22 as not available. Rechecked the Readings and Lecture Notes pages on 2026-10-13: the course has 25 sessions, neither recording nor slides are public for Lec 22 and Lec 25, and the public Lec 20 recording is only 38 minutes long. Per-lecture videos are embedded in each lecture article.
Course and recording entries:
Checked: 2026-10-10.
Content check: verified against the video transcript (2026-10-10): this post embeds no video, so only its statements about the videos were checked: I read the transcripts and descriptions of the 18 lecture videos Lec 01–18 (Lec 19 onward not read) and confirmed that the Fall 2024 lecturers were Isola, Bernstein and Beery (Beery did not only join in 2025); the old "taught by Isola and Bernstein" omitted Beery and was corrected.
Update Log
- 2026-10-10: Added explicit video status and checked recording sources and access notes.
- 2026-10-10: Rechecked video status. Live-checked the official entry: per-lecture recordings are listed there and embedded in each lecture article; the status stays “Official entry or recording index only.”
- 2026-10-10: Checked the video content against its transcript. Corrected the Fall 2024 lecturers (added Sara Beery) and noted who gave which lecture.
- 2026-10-13: Added "OCW Fall 2024, Lecture by Lecture" with links to the new guides for Lectures 2, 6, 19, 20, 21, 23 and 24, and notes on why Lectures 22 and 25 have no public material and why the Lecture 20 recording is only the first half; title scope changed from "Topic Guide" to "Course Overview".
References
- MIT 6.7960 Deep Learning, Fall 2025 course site — source for the 21-lecture schedule, grading, policies, and every public link
- 6.7960 Fall 2025 Project Guidelines PDF — final project requirements, teaming, and compute advice
- PyTorch Tutorial Colab — official Week 2 tutorial notebook
- Foundations of Computer Vision (Torralba, Isola, Freeman) — the free online textbook behind most required readings
- Understanding Deep Learning (Simon Prince) — the free deep learning textbook recommended by the course site
- 6.7960 Fall 2024 course site — the Isola/Bernstein era syllabus and decks
- MIT OCW: 6.7960 Fall 2024 — historical edition with public recordings and full assignments
- 6.s898 Fall 2023 course site — earlier iteration
- 6.390 Intro to ML public notes — the site-recommended ML refresher
- Understanding deep learning requires rethinking generalization (arXiv:1611.03530) — optional reading for Lecture 7
- Language Models are Few-Shot Learners (GPT-3, arXiv:2005.14165) — optional reading for Lecture 11
- Scaling Laws for Neural Language Models (arXiv:2005.10242) — optional reading for Lecture 12
- Training Compute-Optimal LLMs (Chinchilla, arXiv:2203.15556) — optional reading for Lecture 12
- ColBERT (arXiv:2004.12832) — instructor Khattab's signature work, related to Lecture 9
- On this site: Global AI and CS Course Map — definition of the A0–A3 openness tiers
- On this site: MIT AI/ML Course Map — where 6.7960 sits in MIT's curriculum ladder
Loading...