Autoencoders & VAEs
From bottleneck compression to generative models. Ground-up derivations of autoencoders, variational autoencoders, the ELBO, and the reparameterization trick, with interactive MNIST demos.
Weight Initialization
Why random weights need the right scale, and how Xavier and He initialization preserve signal variance through deep networks. Ground-up derivations with interactive demos.
Activation Functions
Why neural networks need nonlinearity, and what happens when you pick sigmoid vs ReLU vs GELU. Interactive demos of each function and its gradient.
Dimensionality Reduction
PCA finds straight lines, t-SNE and UMAP find curves. Interactive demos showing how each method projects high-dimensional data onto 2D.
L1 vs L2 Loss: MAE and MSE
Interactive comparison of L1 and L2 loss functions. Drag outliers around and watch regression lines react differently.
Lasso vs Ridge: L1 and L2 Regularization
Interactive comparison of L1 and L2 regularization. See why Lasso zeros out weights, Ridge just shrinks them, and how the constraint geometry explains it all.
Norms: Measuring Size and Distance
What Lp norms are, why L1 and L2 measure different things, and how the unit ball shape explains everything from Manhattan distance to sparsity.
From MLE to MAP: L2 Regularization is Bayesian in Disguise
Interactive exploration of MLE, MAP estimation, and the Gaussian prior hiding inside weight decay.
DoMINO: How NVIDIA's Physics Surrogate Works
Interactive architecture explorer for NVIDIA's DoMINO model. Walk through every layer of the pipeline that predicts full aerodynamic flow fields from raw STL geometry in seconds.
PCA of SDFs
How many principal components does it take to reconstruct a 3D shape from its signed distance field? Interactive rebuild from the ground up.
What is a signed distance field?
The scalar function that quietly underpins most modern 3D ML. Interactive intuition, minimal math.