Cursarium logoCursarium
advancedFree

Deep Learning

by Yann LeCun & Alfredo Canziani · NYU

4.8
(1,400 reviews)
100K+ enrolled14 weeksUpdated 2024-02

Our Verdict

Worth it — with caveats

NYU's DS-GA 1008 Deep Learning, taught by Turing Award winner Yann LeCun with Alfredo Canziani, is one of the most advanced free deep learning courses on the open web, and it is best taken as a second course after you already know the basics. The full Spring 2021 edition (DLSP21) is released for free as close-captioned lecture videos, written overviews, and executable PyTorch Jupyter notebooks, with no enrollment, login, or certificate. Its distinctive value is LeCun's energy-based-model framing and heavy emphasis on self-supervised learning, transformers, and graph neural networks, content you rarely find in beginner courses. The honest catch: multiple sources note LeCun's lectures are theory-heavy and 'not so student friendly as Andrew Ng's,' and for-credit NYU students on Coursicle criticized the lectures as high-level 'seminar talks' with homework that felt disconnected from the material, so beginners should not start here. As a free, self-paced resource it carries no cost and no risk, and the real binding constraint is your math and prior-ML readiness, not money.

World-class, genuinely advanced free content from a deep learning pioneer, but it is a graduate course with real prerequisites (machine learning + linear algebra) and a polarizing, theory-forward lecture style. It is an excellent 'second' deep learning course for people who already know the fundamentals, and the wrong choice for true beginners, who should start with fast.ai or the DeepLearning.AI specialization first.

Best for: Learners who have already completed an introductory ML course and a first deep learning course and want depth on modern research themes, especially energy-based models, self-supervised / joint-embedding methods, transformers, and graph neural networks. Ideal for graduate students, ML engineers, and researchers comfortable with linear algebra, calculus, backpropagation, and reading PyTorch, who want LeCun's specific conceptual framing and ~50 hours of dense, current material with runnable notebooks.

Skip if: Complete beginners or anyone who has never trained a neural network. Reviewers consistently flag that LeCun's lectures present the big picture rather than step-by-step mechanics, and one NYU student called it the worst class they took at NYU because homework and the group project felt 'unlinked with class material.' People who need hand-holding, a guided coding curriculum, graded feedback, or a completion certificate should pick a structured MOOC instead.

About This Course

NYU's graduate-level deep learning course covering energy-based models, self-supervised learning, and graph neural networks.

What You'll Learn

Neural network fundamentals: gradient descent, the backpropagation algorithm, modules, and training mechanics
Convolutional and recurrent networks (including LSTMs) and the natural-signal properties that motivate convolution
Energy-based models in depth, from foundations to inference and training of latent-variable EBMs and structured prediction
Self-supervised and unsupervised learning, autoencoders, GANs, and contrastive vs. regularized joint-embedding methods
Attention and transformer architectures, plus associative memories
Graph neural networks: graph convolutional networks and graph transformer networks
Optimization for deep learning and applied control topics (planning under uncertainty, the Truck Backer-Upper problem)

Curriculum

Introduction & neural net foundations

History and motivation, gradient descent and backpropagation, neural network inference, modules and architectures, and training. Includes Homework 1 (backpropagation).

Parameter sharing: ConvNets & RNNs

Recurrent and convolutional networks, ConvNets in practice, natural signal properties and convolution, vanilla RNNs and LSTMs. Includes Homework 2 (RNN & CNN).

Energy-based models (foundations)

EBMs Parts I-II, inference and training for latent-variable EBMs, applications, and structured prediction. Includes Homework 3.

Energy-based models (advanced) & unsupervised learning

Advanced EBM concepts, autoencoders (target propagation, PyTorch implementations), GANs, and contrastive vs. regularized joint-embedding methods.

Associative memories: attention & transformers

Attention mechanisms and transformer architectures.

Graphs

Graph transformer networks and graph convolutional networks (Parts I-II).

Control

Planning and control, the Truck Backer-Upper problem, and prediction/planning under uncertainty.

Optimization

Optimization techniques for deep learning (Parts I-II).

Specialized guest topics

Self-supervised learning for vision, low-resource machine translation, speech recognition with graph transformers, and Lagrangian backpropagation.

Prerequisites

  • A prior machine learning course (the for-credit version lists DS-GA 1001 Intro to Data Science and DS-GA 1003 Machine Learning as prerequisites; the free release states DS-GA 1001 or an equivalent ML course)
  • Linear algebra and multivariable calculus
  • Working understanding of gradient descent and backpropagation
  • Python and the ability to read/run PyTorch code in Jupyter notebooks

Instructor

Yann LeCun & Alfredo Canziani

Instructor · NYU

Pros & Cons

Pros

  • Taught by Yann LeCun (Turing Award winner, co-inventor of modern ConvNets) with Alfredo Canziani's well-regarded, approachable practical sessions; you get a research pioneer's actual mental model.
  • Covers genuinely modern, research-grade topics that most courses skip: energy-based models, self-supervised / joint-embedding learning, transformers, and graph neural networks.
  • Completely free and self-paced with no login or paywall: ~50 hours of close-captioned videos, written overviews, and runnable PyTorch Jupyter notebooks, with materials translated by 470+ volunteers into many languages.
  • The DLSP21 edition adds polished motion-graphics visualizations, and the GitHub notebooks let you run real code rather than just watching.

Cons

  • LeCun's lectures are theory-first and high-level; reviewers describe them as 'not so student friendly as Andrew Ng's' and, from for-credit students, as general 'seminar talks' that omit implementation details.
  • Steep prerequisites and pace: one reviewer needed over three months of full-time study, and beginners will likely be lost without prior ML and solid math.
  • No certificate, no graded feedback, and no instructor support for self-learners; for-credit students also reported homework and the group project feeling disconnected from lectures.
  • The free release is the 2020/2021 edition, so the most recent post-2021 advances (e.g., latest large-scale generative/LLM developments) are not covered.

Alternatives To Consider

Frequently Asked Questions

Is Deep Learning free?

Yes — Deep Learning is free to access. Free. All DLSP21 materials (videos, written notes, and PyTorch notebooks) are openly available with no enrollment, login, or payment, and there is no certificate. The for-credit NYU version (DS-GA 1008) is separate, tuition-based, and requires admission to NYU.

Who is Deep Learning for?

Learners who have already completed an introductory ML course and a first deep learning course and want depth on modern research themes, especially energy-based models, self-supervised / joint-embedding methods, transformers, and graph neural networks. Ideal for graduate students, ML engineers, and researchers comfortable with linear algebra, calculus, backpropagation, and reading PyTorch, who want LeCun's specific conceptual framing and ~50 hours of dense, current material with runnable notebooks.

What will you learn in Deep Learning?

Neural network fundamentals: gradient descent, the backpropagation algorithm, modules, and training mechanics; Convolutional and recurrent networks (including LSTMs) and the natural-signal properties that motivate convolution; Energy-based models in depth, from foundations to inference and training of latent-variable EBMs and structured prediction; Self-supervised and unsupervised learning, autoencoders, GANs, and contrastive vs. regularized joint-embedding methods.

What are the prerequisites for Deep Learning?

A prior machine learning course (the for-credit version lists DS-GA 1001 Intro to Data Science and DS-GA 1003 Machine Learning as prerequisites; the free release states DS-GA 1001 or an equivalent ML course); Linear algebra and multivariable calculus; Working understanding of gradient descent and backpropagation; Python and the ability to read/run PyTorch code in Jupyter notebooks.

Is Deep Learning worth it?

World-class, genuinely advanced free content from a deep learning pioneer, but it is a graduate course with real prerequisites (machine learning + linear algebra) and a polarizing, theory-forward lecture style. It is an excellent 'second' deep learning course for people who already know the fundamentals, and the wrong choice for true beginners, who should start with fast.ai or the DeepLearning.AI specialization first.

How we reviewed this course

This is an independent editorial assessment by Cursarium, based on NYU's published course materials and aggregated public learner feedback (last reviewed 2026-06). We have not independently completed the course. Links to providers are standard references, not paid placements.