Every lecture pairs its math with a small, runnable Python script that reproduces the lecture's own worked numerical example — download it, run it, and check the printed numbers against the page. Where a proper implementation is too heavy for a single demo (training a real CNN, a full seq2seq model, a production autoencoder), the card below also links to a stable, official external tutorial or Colab notebook.
💾 24 lectures · all code is plain NumPy unless noted, with a few Keras/TensorFlow examples for the heavier architectures.
One-time environment setup — checks that NumPy, Matplotlib, scikit-learn and TensorFlow are installed before you start the course.
Fits polynomials of increasing degree to a synthetic dataset to visualize underfitting vs. a good fit vs. overfitting.
Trains a single perceptron on the AND-gate dataset from scratch, reproducing the lecture's hand-worked weight updates.
A pure-NumPy forward pass function (X·W1+b1 → ReLU → ·W2+b2 → sigmoid) matching the lecture's worked example.
Plots sigmoid/tanh/ReLU/Leaky ReLU and their derivatives, and reproduces the sigmoid-vs-tanh gradient-update comparison.
Computes binary cross-entropy on a toy prediction and a manual KL-divergence calculation between two discrete distributions.
Reproduces the convergent (η=0.1) and divergent (η=1.1) gradient-descent runs on L(w)=(w−3)², plus a mini-batch GD demo.
Full from-scratch backprop implementation reproducing the lecture's exact worked forward+backward pass and a 200-step training loop.
Computes one update step under plain SGD, Momentum, AdaGrad and Adam from the same starting point, matching the lecture's comparison.
Reproduces every metric worked example on the page — MSE/MAE, two confusion matrices, and the full 21-row ROC/PR threshold scan.
The same architecture built two ways: a hand-derived NumPy training loop, and the equivalent Keras/TensorFlow model.
Reproduces the MLP parameter-explosion numbers, the conv filter-bank parameter formula, and the 1×1-conv channel-reduction example.
Implements the output-size formula, the exact 4×4 max-pooling example, and the Sobel edge-detection worked example.
A complete, runnable Keras CNN for MNIST (Conv→Pool→Conv→Pool→Flatten→Dense→Softmax), the instructor's own layer sizes.
A scalar RNN forward pass reproducing h₁, h₂, h₃ from the lecture, showing the same weights reused at every timestep.
Multiplies chains of <1 and >1 gradient factors to show, numerically, how a long BPTT chain vanishes or explodes.
One full LSTM cell timestep in NumPy — forget/input/candidate/output gates — reproducing the lecture's exact numeric walkthrough.
A Keras Bidirectional(LSTM) layer plus an illustrative encoder-decoder skeleton showing how the pieces connect.
Byte Pair Encoding merge-training implemented from scratch on the instructor's own corpus (huggingface, hugging, face, hug, …).
A complete, seeded NumPy implementation of scaled dot-product self-attention — Q/K/V projection, scaled softmax, weighted-sum output.
Reproduces the competitive-learning winner-update example and a from-scratch K-means loop on a small synthetic 2D dataset.
One full Contrastive Divergence (CD-1) training step on a tiny 2-visible/1-hidden RBM, reproducing every number in the lecture exactly.
A linear autoencoder forward pass (clean + masked-noise denoising), reproducing MSE≈0.1444/0.1921, plus a KL sparsity-penalty snippet.
Computes the contractive penalty (Jacobian Frobenius norm, reproduces 0.29) plus a Keras Conv2D/Conv2DTranspose autoencoder skeleton.