Interactive, browser-based demos for learning the fundamentals of computer vision.
| Demo | What it teaches | Link |
|---|---|---|
| Image Normalization | How pixel values are rescaled before training, and how normalization layers work inside networks | ▶ |
| ⛰️ Optimizers | How SGD, Adam and the newest optimizers (Lion, Muon, SOAP and more) find the lowest loss | ▶ |
| 〰️ Activation Functions | How ReLU, GELU, SwiGLU and others shape a network, with a tiny network you can train live | ▶ |
| 🔲 Convolution and Filters | How a small grid of weights blurs, sharpens and finds edges, with the full sum shown for any pixel | ▶ |
| 🎨 Color Spaces | RGB, HSV, LAB and YCbCr channels, and why picking out a color is easier in HSV | ▶ |
| 🔍 Resizing and Interpolation | Nearest, bilinear, bicubic and area resizing, and how shrinking creates aliasing | ▶ |
| 🧱 CNN Layer Explorer | Kernel size, stride, padding and dilation, output sizes, parameters and receptive fields | ▶ |
| 🗜️ Pooling and Downsampling | Max, average and strided pooling, and why one-pixel shifts can change a CNN's output | ▶ |
| 🎬 Classification & Regression Basics | How a network's final layer, activation and loss function change across binary classification, multi-class classification, and regression | ▶ |
See what normalization does to an image, one pixel at a time. Pick a photo, pick a method, then point at any pixel to see the exact arithmetic behind the result.
Pixel normalization methods
| Method | What it does |
|---|---|
| Scale to 0–1 | Divides every value by 255, so the numbers change but the picture doesn't |
| Min–max | Stretches the darkest value to 0 and the brightest to 1 |
| Percentile clip | Like min–max, but ignores extreme outliers first |
| Z-score | Centers values on 0 with a spread of 1 |
| ImageNet mean/std | The fixed normalization pretrained models expect |
| Histogram equalization | Spreads pixels evenly across all brightness levels |
Normalization layers in neural networks
Batch norm, layer norm, instance norm and group norm, shown on a batch of 4 images so you can see exactly which values get averaged together.
- Pixel inspector: point at or tap any pixel to see the before and after values with the full calculation.
- Live histograms and statistics: watch the min, max, average and spread change for each color channel.
- Three explanation levels: switch between new to this, comfortable with math and training models.
- Problem images: fog, a dark night scene, a color cast and a well-exposed photo show where each method helps and where it fails.
- Use your own photo: it's processed in your browser and never uploaded anywhere.
Watch optimizers race downhill on the same loss landscape, then see why the newest ones treat weights as matrices instead of separate numbers.
Optimizers in the 2-D race
| Optimizer | Year | Big idea |
|---|---|---|
| SGD | 1951 | Step straight downhill |
| Momentum | 1964 | Build up speed in consistent directions |
| Nesterov | 1983 | Look ahead before stepping, to overshoot less |
| AdaGrad | 2011 | A separate, shrinking step size for each direction |
| RMSprop | 2012 | Step sizes based on recent gradients only |
| Adam | 2014 | Momentum plus per-direction step sizes |
| AdamW | 2017 | Adam with weight decay done right |
| Lion | 2023 | Step by the sign of the momentum only |
| Sophia | 2023 | Uses curvature, with clipped steps |
| Schedule-Free | 2024 | Averaging replaces learning-rate decay |
| ADOPT | 2024 | Reordered Adam that always converges |
| AdEMAMix | 2024 | Mixes a fast and a very slow memory |
| Cautious | 2024 | Skips directions where momentum disagrees with the gradient |
Matrix-aware optimizers
Muon, Shampoo and SOAP, shown on a 16×16 weight matrix to explain why they keep learning when SGD and Adam stall.
- Four terrains: a narrow valley, a curved "banana" valley, a saddle point and a bumpy surface with many local minima.
- Live optimizer internals: watch the momentum, step-size memory and every update as numbers, step by step.
- Controls to experiment: change the learning rate, add noisy gradients like real mini-batches, or click the map to pick a new starting point.
- Three explanation levels: the same new to this, comfortable with math and training models switch as the normalization demo.
- What the research says: a table of optimizers published from 2023 to 2026, plus links to the original papers.
See how the small curves inside a network let it learn anything more than straight lines. Compare their shapes, train a tiny network with each one, and find out why modern models use the ones they do.
Activations in the explorer
| Activation | Year | Big idea |
|---|---|---|
| Step | 1943 | Fire or don't; can't be trained with gradients |
| Sigmoid | 1980s | A smooth step from 0 to 1 |
| Tanh | 1990s | Like sigmoid, but centered on zero |
| Softplus | 2001 | A smooth, always-positive ReLU |
| ReLU | 2010 | Keep positives, zero out negatives |
| Leaky ReLU | 2013 | ReLU that lets a little through for negatives |
| ELU | 2015 | Negatives curve smoothly to a floor |
| GELU | 2016 | Smooth ReLU, used in BERT, GPT and ViT |
| SELU | 2017 | Keeps signals a steady size on its own |
| Swish / SiLU | 2017 | x times sigmoid(x), found by automated search |
| Mish | 2019 | A smooth activation popularized by YOLOv4 |
| Hard-Swish | 2019 | A cheap Swish built for phones |
| Squared ReLU | 2021 | ReLU squared, for sparse and efficient models |
Beyond single curves
- Softmax: how the last layer of a classifier turns raw scores into probabilities.
- Gated activations: GLU, ReGLU, GeGLU and SwiGLU, the building block of today's large models, and why SwiGLU's fast growth led to newer fixes.
- Function explorer: overlay any activations, show their slopes, read exact values, and adjust parameters like Leaky ReLU's α or Swish's β.
- Train a tiny network live: watch it separate circle, XOR, two-moons and spiral data with each activation, including dead ReLU neurons piling up.
- Vanishing gradients: measure how the error signal shrinks or explodes through a network up to 60 layers deep.
- Softmax playground: move class scores and the temperature to see how the probabilities change.
- Three explanation levels: the same new to this, comfortable with math and training models switch as the other demos.
Type your own text or drag a few sliders, and watch the same three problem types a neural network commonly solves respond live: sorting into two classes, sorting into many classes, and predicting a plain number.
Three problems, one shared pipeline
| Problem type | Example task | Output layer |
|---|---|---|
| Binary classification | Movie review → positive or negative | 1 unit, sigmoid |
| Multi-class classification | Newswire → 1 of 46 topics | N units, softmax |
| Scalar regression | House features → a price | 1 unit, no activation |
Along the way
- Why repeated words don't add anything in a multi-hot vector — presence, not frequency.
- Why a hidden layer narrower than the number of output classes becomes an information bottleneck.
- Why a linear output layer can predict any real number, while sigmoid and softmax both stay capped.
- Why K-fold validation exists: with too little data, a single train/validation split is too noisy to trust.
- Live text classifier: type a movie review and watch its tokens, multi-hot vector, and sigmoid score update as you type.
- Architecture builder: change the number of hidden layers and units, and watch the network diagram, parameter count, and overfitting speed respond.
- Newswire topic sorter: type a news snippet and see softmax probability bars across 6 real Reuters topic categories.
- Price estimator: four sliders, already in normalized units, driving a live linear prediction.
- K-fold selector: click through 4 validation rounds to see which slice of data is held out each time.
Each demo is a single self-contained HTML file. To use it offline, download it and open it in any modern browser.