NeuralNetVis: A 3D Interactive Platform for Visual Exploration of Neural Networks in Browsers

NeuralNetVis is an open-source neural network visualization tool built with TensorFlow.js and Three.js, aiming to break the "black box" barrier of neural networks. This platform supports real-time training, 3D network structure display, and multi-dataset experiments in browsers, making deep learning principles intuitive and tangible. The project is maintained by Daniel Marcus; the online demo is available at https://neuralnetvis.app, and the GitHub repository link is https://github.com/daniel-marcus/neuralnetvis.

Neural networks have long been regarded as "black boxes", with the intermediate processes between input and output being difficult to understand intuitively. As a browser-based interactive 3D platform, NeuralNetVis helps users observe and explore the working mechanisms of neural networks through visualization. Users can build and train models by themselves, and view dynamic processes such as weight updates and activation propagation without installing any software.

NeuralNetVis uses TensorFlow.js for model construction, training, and inference, and Three.js for 3D rendering. Key highlights include GPU direct data transfer supported by WebGPU technology, which improves rendering efficiency and interaction smoothness. The tech stack also includes Next.js App Router, Zustand state management, and WebAssembly backend for accelerated computing. The best experience is achieved on the desktop version of Chrome browser.

The platform has built-in multiple types of datasets: classification tasks include MNIST, Fashion MNIST, CIFAR-10/100, etc.; regression tasks include Auto MPG, California Housing; NLP tasks include IMDb sentiment analysis. Pre-trained models include MediaPipe hand key point detection and MobileNetV2 image classification. Advanced users can import custom LayersModel from Keras3.x (.keras format) or Keras2.x (converted via tfjs-converter).

The target audience includes deep learning beginners (to understand core concepts), educators (for classroom demonstrations), researchers (to verify model architectures), and developers (to learn Web ML practices). Typical use cases include browser prototype design, science popularization for non-technical personnel, model debugging, and teaching material generation. Reference resources include Stanford CS231n, Adam W. Harley's CNN visualization research, TensorFlow Playground, etc.

Current limitations: limited mobile support, dependence on WebGPU (requires modern browsers like Chrome), and only supports LayersModel import. In the future, as WebGPU becomes more popular and the TensorFlow.js ecosystem matures, such tools will play a greater role in AI education and popularization.

NeuralNetVis transforms abstract algorithms into interactive visual experiences, providing an excellent starting point for deep learning learners to break through the "black box" dilemma. It is recommended that interested users visit https://neuralnetvis.app to experience it personally, or check the GitHub repository for more details.