Deconstructing CNNs from Scratch
Explore the mathematics of convolutions, weight sharing, and hierarchical feature learning — built entirely from scratch in pure PyTorch.
Part 1
The Math of Convolutions
Breaking down the 2D convolution operation, local connectivity, weight sharing, pooling, and the inductive biases that make CNNs powerful for visual data.
Part 2
PyTorch Implementation
Building Conv2D, pooling layers, and complete CNN architectures (LeNet-5,
SimpleCNN, DeepCNN) entirely from scratch in PyTorch.
View Code on GitHub
Part 3
Training & Visualizing Features
Training on MNIST, visualizing learned convolutional filters, and analyzing hierarchical feature map activations to peer inside the black box.