Northeastern University · CS 7150

Deep Learning class demos

Interactive visualizations and Jupyter notebooks used in lecture. Everything here lives in github.com/CS7150/classdemos — notebooks open straight into Colab, and the interactive pages run in the browser with nothing to install.

Interactive demos

Runs in your browser

Rosenblatt's perceptron

The update rule, the margin, and the mistake bound, shown side by side in data space and weight space. Step through a run, scrub forwards and back, or watch the worst case walk the corners of the region the proof still permits.

Open the demo →

Guides

Annotated transcript

Using an AI agent to learn the perceptron algorithm

An AI agent can give you the convergence proof in a single prompt. Getting it to explain the proof so you actually understand it took a few hours, and this is what that looked like: twenty rounds of saying “I don’t understand”, with the figure at each stage and the reasoning behind each next question in the margin.

Read the transcript →

Notebooks

Single neurons and what they can separate

The perceptron, its learning rule, and the limits of one linear threshold unit.

Perceptron Learning in 2D Rosenblatt's algorithm on linearly separable 2D data, visualized step by step.

Colab Source
Fraction Linearly Separable How quickly the share of boolean functions computable by one neuron collapses as inputs grow.

Colab Source
Logic Networks McCulloch–Pitts neurons with a sign nonlinearity, wired into logic circuits.

Colab Source

XOR and the first hidden layer

What one layer cannot do, and what happens once you train two or more.

Learning XOR A two-layer tanh network trained on the XOR truth table by hand-rolled gradient descent.

Colab Source
Learning XOR (batch variant) The same network trained on the full truth table each step instead of one row at a time.

Colab Source
Learning Four-Layer XOR A deeper net on XOR, checkerboard, and sine-wave targets, with a widget for the training run.

Colab Source
Rumelhart, Hinton & Williams Mirror Reproduces the 1986 symmetry-detection experiment on all 64 six-bit patterns.

Colab Source

Training dynamics

Watching decision boundaries and loss surfaces move while a small network learns.

Two-Layer 2D Net Widget An interactive two-layer ReLU classifier: hidden-unit lines and the boundary they compose.

Colab Source
Explore Two-Layer 2D Net The same network with the training run scrubbable, for comparing optimizers and targets.

Colab Source
2D Optimization Gradient descent on a small 2D classification problem, plotted alongside the decision boundary.

Colab Source

Backpropagation and loss functions

Derivatives by hand, and what different losses actually reward.

By-Hand Backpropagation Build a computation graph from scratch and implement the forward and backward passes yourself.

Colab Source
Graph KL Loss KL divergence, cross-entropy, MSE and L1 over the same softmax, graphed against each other.

Colab Source
Softmax Chart The figure for turning a pair of logits into probabilities and comparing them to a target.

Colab Source

Convolution and vision

Convolution as a matrix, and the fully-connected versus convolutional comparison.

Toeplitz Convolutions 1D and 2D convolution written as Toeplitz matrices: shift equivariance, padding, parameter counts.

Colab Source
CIFAR-10 Classifier Homework notebook: a fully-connected baseline on CIFAR-10, then a convolutional network.

Colab Source

Large pretrained models

Poking at CLIP and Stable Diffusion, and building the sample data they run on.

CLIP Doctor Rank images by cosine similarity to text prompts using a pretrained CLIP ViT-B/32.

Colab Source
Diffusion Take a Stable Diffusion pipeline apart and look at the networks inside it.

Colab Source
Data Set Maker How the coco_humans sample set used by the other notebooks was built from COCO 2017.

Colab Source