1 · ANN foundations
From the single perceptron to backpropagation and the vanishing-gradient barrier — the core machinery every deep net uses.
●Track 9 · live
Deep Learning
ANNs, CNNs, RNNs, LSTMs, Transformers — animated end-to-end.
From the perceptron to multi-head attention. PyTorch/TensorFlow code is read-along (they don't run in Pyodide), but every key idea — backprop, convolution stride, attention heatmaps, transformer blocks — gets a cinematic scene that makes the math click.
49 / 49 lessons10 chapters
2 · Activations, loss, optimizers
Sigmoid to GELU, MSE to cross-entropy, SGD to Adam — the three knobs you turn to make a deep net train.
3 · Training tricks + projects
Weight init, dropout, batch norm — plus two complete projects: ANN binary classifier and ANN regressor, hand-rolled in numpy with PyTorch read-alongs.
4 · Convolutional Neural Networks
From the convolution operation to a complete image classifier — the architecture that ruled computer vision for a decade.
- ›15 · Why CNNs: spatial structure + parameter sharing~12 min
- ›16 · Convolution: kernels, strides, and feature maps~14 min
- ›17 · Padding + pooling: controlling spatial dimensions~12 min
- ›18 · Channels: RGB images and multi-filter convolutions~12 min
- ›19 · Complete CNN architecture: Conv → BN → ReLU → Pool~14 min
5 · Recurrent Neural Networks
From RNN intuition through BPTT and the vanishing-gradient problem, to an end-to-end sentiment classifier.
6 · LSTM, GRU, Bidirectional
Gated recurrent architectures that solved the vanishing-gradient problem and ruled NLP from 2014 to 2017.
7 · Seq2Seq + Attention
Encoder-decoder, the bottleneck problem, attention as soft lookup, and self-attention — the bridge from RNNs to Transformers.
8 · Transformers
Multi-head attention, positional encoding, the encoder/decoder block, and a from-scratch mini-GPT — the architecture behind every modern LLM.
9 · Pretrained models & LLMs
Subword tokenization, self-supervised pretraining, the three architecture families, transfer learning, LoRA, and instruction-tuning + RLHF — how real LLMs are built, adapted, and aligned.
10 · Generative models
Autoencoders, VAEs, GANs, diffusion (DDPM), latent/conditional diffusion (Stable Diffusion), and how the generative families compare — the creative side of deep learning.