CNN visualizer
See what a convolutional network looks for - its filters, its feature maps and where it looks - layer by layer.
Inside a CNN
A CNN slides small filters over the picture; each lights up where it finds its pattern - an edge, a curve, a corner. Block by block, simple patterns combine into parts of objects.
NeuroScope shows every layer's feature maps, the patch of the picture that excites each detector most, which parts of the image the answer depends on, and one convolution worked out number by number.
Train one yourself
Start with handwritten digits, then try Fashion-MNIST, chest X-rays, blood cells or your own ZIP of images. Training runs on your graphics card in the browser.
Or load a famous pretrained network - ResNet, MobileNet, SqueezeNet - and look inside it with your own photo.
Try it in a minute
- Press Try it now - digits and a small CNN are set up.
- Train it, then open Look inside on any digit.
- Hover the feature maps and detectors; open the exact convolution.
Questions
Does training run on my GPU?
Yes - in the browser through WebGPU or WebGL. Without a usable graphics card, it can run on the NeuroScope server.
Can I use pretrained models?
Yes: ResNet-18/34, MobileNet v2/v3 and SqueezeNet run in the browser; bigger ones on the server.
What is Grad-CAM?
A map of which regions of the image pushed the network towards its answer.