d2l.ai // rebuilt as an interactive lab
Deep learning you can scroll, drag, and break.
Every chapter of Dive into Deep Learning as a visual story: algorithms running live in your browser, figures that advance as you scroll, and labs where you turn the dials yourself.
Foundations
Introduction→
Why you can't hand-write the rule, how a program with knobs learns it from data instead, and why deep learning works now — the whole book in one page.
4 scroll acts · Knob trainer · Death-cap decision · Köbel's mean
Preliminaries→
Tensors, broadcasting, calculus, automatic differentiation, and probability — the toolkit everything else builds on.
5 scroll acts · Broadcast explorer · Coin convergence · Bayes test
Linear Regression→
The simplest model that learns: fit a line to data, descend a loss surface step by step, and watch minibatch SGD wobble its way to the optimum.
3 scroll acts · Fit-it-yourself lab · SGD trainer
Softmax Classification→
From predicting numbers to predicting categories: logits, softmax, cross-entropy, and a real image classifier.
4 scroll acts · Softmax probe · Temperature dial · Test-set size
Multilayer Perceptrons→
Stack layers, add nonlinearity, and suddenly the model can learn anything — plus backprop, dropout, and why deep nets are trainable at all.
4 scroll acts · Boundary bender · Activation compare · Dropout sampler
Architectures
Builders' Guide→
Layers, modules, parameters, GPUs — the software engineering of deep learning, running live in your browser.
4 scroll acts · Block builder · Initializer lab · Centered layer
Convolutional Neural Networks→
Slide a kernel over an image and watch features appear: convolution, padding, stride, channels, pooling, and LeNet.
4 scroll acts · Kernel explorer · Size dials · Pooling invariance
Modern CNNs→
AlexNet to ResNet to DenseNet — the architecture ideas that made deep learning deep, plus the two that outlived vision: batch norm and residual connections.
5 scroll acts · Batch-norm dials · Residual depth · Architecture explorer
Recurrent Neural Networks→
Models with memory: unroll a cell through time, train a character-level language model, and meet vanishing gradients.
4 scroll acts · Autoregression lab · Char-RNN sampler · Gradient flow
Modern RNNs→
LSTMs, GRUs, encoder–decoder, and sequence-to-sequence translation with beam search.
5 scroll acts · Memory carousel · Beam search · BLEU scorer
Attention & Transformers→
Queries, keys, and values; a scoring function that scales safely; heads that specialize; positions as waves; and the full Transformer — trained live to reverse sequences and read images.
6 scroll acts · Pooling explorer · Positional codes · Reverse-it probe
Training at Scale
Optimization→
Race SGD, momentum, AdaGrad, RMSProp, AdaDelta, and Adam across pathological loss landscapes — then schedule the learning rate like it's 2015 ImageNet.
6 scroll acts · Optimizer race · Momentum region · LR scheduler
Computational Performance→
Why the same model trains in a week or three months: compilation, the async queue behind every GPU call, the hardware underneath, and data parallelism across many GPUs.
6 scroll acts · Pipeline driver · Latency ladder · Ring allreduce
Applications
Computer Vision
Augmentation, fine-tuning, object detection, segmentation, and neural style transfer.
NLP: Pretraining
word2vec, GloVe, subword embeddings, and BERT — words as vectors you can do arithmetic on.
NLP: Applications
Sentiment analysis and natural language inference with RNNs, CNNs, and fine-tuned BERT.
Recommender Systems
Matrix factorization to deep factorization machines on real ratings data.
Frontiers
Reinforcement Learning
Markov decision processes, value iteration sweeping a gridworld, and Q-learning.
Gaussian Processes
Priors over functions: uncertainty you can see.
Hyperparameter Optimization
Random search, successive halving, and asynchronous scheduling.
Generative Adversarial Networks
A generator and a discriminator locked in a duel — watch fake data sharpen into the real thing.
Bonus
Math for Deep Learning
Geometry of linear algebra, eigendecompositions, calculus, probability distributions, and information theory.