Neural Network X-Ray

Draw a character, watch 13 CNN layers fire. Interactive CNN visualiser running entirely in the browser through ONNX Runtime on WebAssembly. Latin and Bengali scripts, 146 classes, about 980K training images.

Live GitHub

FIG. 0Neural Network X-Ray intro, 21 seconds, with sound

Neural Network X-Ray screenshot — **FIG. 1**Neural Network X-Ray, live at ocr.soyebjim.me

**DETAIL A**All 64 pooled channels of the third convolution

**DETAIL B**Softmax verdict with the runner-up classes

Classes146
Training images~980K
Epochs40
Parameters1.7M, a 7 MB ONNX file
Validation accuracy90.7%
stackNext.js 16, React 19, TypeScript, ONNX Runtime, WebAssembly, PyTorch
linksLive, GitHub, Model weights, Dataset

You draw a letter or digit on a canvas and every layer of the network lights up as it processes the stroke: convolution, ReLU, pooling, the dense layers and the final softmax. Inference runs in the browser with ONNX Runtime Web, so nothing is uploaded.

The model was trained on EMNIST ByMerge merged with BanglaLekha-Isolated: 146 classes covering A to Z, a to z, 0 to 9 and Bengali ka to ha plus compound characters, from roughly 980,000 images over 40 epochs. Weights and the merged dataset are published on Hugging Face.

Built with Next.js 16, React 19 and TypeScript.

DRAWN BYS. P. Jim

SHEET5.2 of 8

REV2026-09

TypeOpen source

Neural Network X-Ray