Deep Learning and Generative Models: Representations and Self-Supervised Learning
Learn how neural networks form internal representations, how those representations transfer to new tasks, and how useful features can be learned without labels. Part four in a five-part series, this online course in the Deep Learning and Generative Models program covers representation analysis, autoencoders, metric learning, and self-supervised contrastive learning. These methods help you work with data when human labeling is costly or limited. You will learn to evaluate representations and measure similarity between examples—skills that support applications such as finding related documents, grouping similar images, and adapting learned features to new prediction tasks.
Course Information
Certificate Track
Learn for Free
About this Course
The value of a trained network extends beyond its final predictions. Representation learning uses features inside the model to reveal structure in data and support new tasks. Examine what hidden units respond to, compare representations across networks, and use k-means clustering and principal component analysis (PCA) as classical baselines.
Reconstruction and similarity offer two routes to useful representations. Autoencoders and masked autoencoders rebuild missing or compressed inputs, while deep metric learning and self-supervised contrastive learning bring related examples together and separate dissimilar ones. Positive pairs and data augmentation determine what a model preserves or ignores.
To determine which representation is useful, train an autoencoder and a contrastive encoder on the same data in PyTorch. Freeze both encoders, evaluate their embeddings with the same downstream classifier and data split, and select the representation that best supports a stated task.
This is the fourth of five courses in the Deep Learning and Generative Models Program:
- Neural Network Foundations with PyTorch
- Training, Inference, and Model Evaluation
- Neural Network Architectures
- Representations and Self-Supervised Learning
- Diffusion and Autoregressive Models
What you'll learn
- Analyze what a trained network encodes in its hidden representations
- Use k-means and principal component analysis as classical baselines
- Train autoencoders and masked autoencoders using reconstruction objectives
- Train metric-learning and contrastive encoders
- Explain how data augmentations shape invariance in learned representations
- Evaluate frozen embeddings through controlled downstream probes
How you'll learn
Real-World LearningLearn from MIT faculty and experts who ground their teaching in real-world cases rather than mathematical models, making the material approachable for all.
Practical ApplicationApply your new knowledge with hands-on, practical exercises drawn from healthcare, sports, finance, sustainability, and more.
- Learn On Demand
Access all course content online with complete flexibility to study at your own pace.
AI-Enabled SupportDeepen your understanding of the course material and get help on assignments from AskTIM, the AI assistant built by MIT researchers.
Stackable CredentialsEarn an MIT Open Learning certificate at each milestone—module, course, and program—demonstrating your AI expertise. Available in paid courses only.
Prerequisites
This program is designed for:
- Aspiring data scientists and machine learning engineers who want a structured introduction to deep learning
- Quantitative researchers who want to apply deep learning to technical or scientific problems
- Working professionals who need to evaluate, adapt, or deploy modern machine learning models
- Advanced undergraduate and graduate students in computer science, data science, statistics, mathematics, engineering, and related quantitative fields
- Continuing education learners seeking hands-on experience with deep learning and generative models
You should be comfortable programming in Python and working with linear algebra, multivariable calculus, including partial derivatives and the chain rule, and undergraduate probability and statistics. The Probability and Statistical Data Analysis series or equivalent coursework provides suitable preparation. No prior deep learning experience is required.
The labs run in Google Colab and are designed for its free GPU tier, so no local installation is required.
Meet your instructors
Kaiming He
Associate Professor, MIT Electrical Engineering and Computer Science
Dr. Kaiming He is a tenured associate professor in MIT’s Department of Electrical Engineering and Computer Science and a part-time Distinguished Scientist at Google DeepMind. He earned his PhD at the Chinese University of Hong Kong and his bachelor’s degree at Tsinghua University. His research focuses on computer vision and deep learning, particularly models that learn representations of the world, with the long-term goal of augmenting human intelligence. He is known for his contributions to deep residual networks (ResNets), whose residual connections are widely used in modern neural networks. His work also includes Faster R-CNN and Mask R-CNN for visual recognition, and MoCo and masked autoencoders for self-supervised learning. Before joining MIT in 2024, he held research positions at Microsoft Research Asia and Facebook AI Research. His honors include best paper awards at CVPR and ICCV, the PAMI Young Researcher Award, and Test of Time awards at ICCV, NeurIPS, and CVPR. He served as a program chair for ICCV 2023 and has taught MIT courses in deep learning, deep generative models, and computer vision. Faculty website, MIT Engineering profile
