Learn Deep Learning

Build neural networks from scratch, train them, and understand the architectures behind modern AI.

Start learning 17 lessons · about 8.5 hours · free
About this track

Deep learning powers image recognition, speech, translation and language models like Claude. In this track you’ll build it from the ground up - not by calling a library, but by writing every piece yourself in numpy, so nothing is magic.

You’ll help Neuro, the robot at Doodle Lab, learn to see. You’ll work with tensors and broadcasting, build neurons, layers and a forward pass, choose loss functions, derive backpropagation and write your own autograd engine, compare SGD, momentum and Adam, and train a network end to end. Then you’ll make training work in practice - initialization, normalization, dropout, weight decay, debugging loss curves - and meet the architectures that matter: convolutional networks, recurrent networks and transformers with attention, before seeing how it all maps onto PyTorch. Along the way you’ll step through backprop node by node, probe activation functions, slide convolution kernels and watch attention. Every exercise runs Python in your browser, and the capstone trains a doodle classifier from scratch.

Before you start
  • Comfortable writing basic Python (loops, functions, lists).
  • No deep learning experience needed; numpy is introduced as you go.
  • School algebra; the calculus you need (derivatives and the chain rule) is explained when it’s used.
  1. Unit 1 · 0/4 lessons

    Neural network foundations

    Tensors and shapes, neurons and activations, and the forward pass of a multilayer network. Badge: Layer builder

    1. 0x001What is deep learning?Meet layered neural networks, see why depth matters, and find where deep learning shines. 20 min 0/2 exercises solved 20 min 0/2 exercises solved
    2. 0x102Tensors, shapes and broadcastingStore data as multi-dimensional arrays, reason about their shapes, and let broadcasting do the loops. 26 min 0/2 exercises solved 26 min 0/2 exercises solved
    3. 0x203Neurons and activation functionsBuild an artificial neuron, explore activation functions, and see why nonlinearity is essential. 26 min 0/2 exercises solved 26 min 0/2 exercises solved
    4. 0x304Layers and the forward passStack layers into a multilayer perceptron, run a batch through it, and count its parameters. 24 min 0/2 exercises solved 24 min 0/2 exercises solved
  2. Unit 2 · 0/5 lessons

    How networks learn

    Loss functions, backpropagation, your own autograd engine, optimizers and a full training loop. Badge: Gradient follower

    1. 0x405Loss functionsMeasure how wrong a network is with mean squared error and cross-entropy, computed stably from logits. 24 min 0/2 exercises solved 24 min 0/2 exercises solved
    2. 0x506BackpropagationCompute every gradient in a network with one backward sweep of the chain rule. 32 min 0/2 exercises solved 32 min 0/2 exercises solved
    3. 0x607Build a tiny autograd engineWrite a Value class that records every operation and backpropagates automatically - the core idea behind PyTorch. 32 min 0/2 exercises solved 32 min 0/2 exercises solved
    4. 0x708Gradient descent and optimizersTurn gradients into weight updates with SGD, momentum and Adam, and tune the learning rate. 28 min 0/2 exercises solved 28 min 0/2 exercises solved
    5. 0x809Train a network end to endPut forward pass, loss, backprop and updates together and train a network from scratch. 34 min 0/2 exercises solved 34 min 0/2 exercises solved
  3. Unit 3 · 0/3 lessons

    Making training work

    Initialization, normalization, residual connections, regularization and debugging. Badge: Training whisperer

    1. 0x9010Initialization, normalization and residualsKeep signals and gradients healthy through deep networks with good initialization, batch and layer norm, and skip connections. 30 min 0/2 exercises solved 30 min 0/2 exercises solved
    2. 0xA011Overfitting and regularizationSpot overfitting and fight it with more data, weight decay, dropout, augmentation and early stopping. 28 min 0/2 exercises solved 28 min 0/2 exercises solved
    3. 0xB012Debugging trainingRead loss curves, run the sanity checks experts use, and fix exploding gradients and NaNs. 26 min 0/2 exercises solved 26 min 0/2 exercises solved
  4. Unit 4 · 0/4 lessons

    Architectures

    Convolutional networks for images, recurrent networks for sequences, transformers with attention, and PyTorch. Badge: Network architect

    1. 0xC013Convolutional networksSlide small learned filters over images, pool the results, and build the CNNs that taught computers to see. 32 min 0/2 exercises solved 32 min 0/2 exercises solved
    2. 0xD014Recurrent networksProcess sequences with a hidden state carried through time, and see why LSTMs and GRUs were invented. 28 min 0/2 exercises solved 28 min 0/2 exercises solved
    3. 0xE015Attention and transformersLet every position look at every other with attention, and assemble the transformer block behind modern AI. 34 min 0/2 exercises solved 34 min 0/2 exercises solved
    4. 0xF016Deep learning with PyTorchMap everything you built by hand onto PyTorch: tensors, autograd, modules, optimizers, data loaders and fine-tuning. 30 min 0/2 exercises solved 30 min 0/2 exercises solved
  5. Unit 5 · 0/1 lessons

    Capstone

    Build, train, debug and evaluate a doodle classifier from scratch. Badge: Doodle decoder

    1. 0x10017Capstone: teach Neuro to read doodlesBuild, train, debug and evaluate a doodle classifier from scratch in numpy - every piece of the track in one project. 50 min 0/3 exercises solved 50 min 0/3 exercises solved
Gostou da aula? 😆👍
Apoie nosso trabalho com uma doação:
Quero contribuir