The canonical building blocks of modern deep learning — from basic representations and sequence models to advanced generative frameworks.
A deep dive deconstructing Autoencoders from first principles. From the mathematics of information bottlenecks and the manifold hypothesis, to building four variants (Vanilla, Denoising, Sparse, Convolutional) from scratch in PyTorch, to visualizing what the bottleneck learns about data structure.
A deep dive deconstructing Generative Adversarial Networks. From the minimax game, Jensen-Shannon divergence, and Nash equilibrium theory, to building both a Vanilla GAN and DCGAN from scratch in PyTorch, to watching two competing networks learn to generate convincing digits from pure noise.
A deep dive deconstructing Variational Autoencoders. From the ELBO derivation and KL divergence, to the reparameterization trick that enables backpropagation through stochastic sampling, to walking through learned latent spaces and interpolating between digits.
A deep dive deconstructing Convolutional Neural Networks. From the mathematics of 2D convolutions and weight sharing, to building Conv2D layers, pooling operations, and complete architectures (LeNet-5, SimpleCNN, DeepCNN) from scratch in PyTorch, to visualizing learned filters and hierarchical feature maps on MNIST.
A deep dive deconstructing Recurrent Neural Networks from first principles. Covering the mathematics of the recurrence relation, a pure PyTorch sequence architecture, and visualizing bounded hidden state integration over time.
A deep dive deconstructing Long Short-Term Memory networks. From the mathematics of gated recurrence and the Constant Error Carousel, to building a multi-layer pure PyTorch architecture, to analyzing gate dynamics on long-range dependencies.
A deep dive deconstructing Residual Networks. From the mathematics of identity skip connections and gradient highways, to building ResidualBlocks, BottleneckBlocks, and the full ResNet-18/34/50/101/152 family from scratch in PyTorch, to visualizing activation flow and learned feature maps on CIFAR-10.
A deep dive deconstructing Transformers from first principles. From the mathematics of Multi-Head Self-Attention and Positional Encodings, to a pure PyTorch sequence-to-sequence implementation, to visualizing dynamic routing on a sequence reversal task.
A deep dive deconstructing Graph Neural Networks. From the mathematics of message passing and neighborhood aggregation, to building GCN and GAT layers from scratch in PyTorch, to classifying nodes in Zachary's Karate Club using only graph topology and 4 labeled examples.
A deep dive deconstructing Vision Transformers. From the mathematics of patch embeddings and positional encodings, to building a complete ViT with CLS token classification from scratch in PyTorch, to visualizing attention maps and confirming that ViTs need large data to outperform CNNs.
A deep dive deconstructing Capsule Networks. From the mathematics of equivariance and dynamic routing-by-agreement, to building PrimaryCapsules, DigitCapsules, and a reconstruction decoder from scratch in PyTorch, to demonstrating superior rotation robustness over standard CNNs.