Neural Network Library in TypeScript

🚀 LIVE DEMO of digit recognition using React.

Overview

This project provides a comprehensive neural network library built in TypeScript with practical applications for image recognition. The library includes:

• Core Library (lib/) - Neural network implementation with various activation and loss functions
• Node.js Application (example/node/) - Command-line tool for training and recognition
• React Application (example/react/) - Interactive web interface for digit recognition

Features

Neural Network Library

• Multiple activation functions (ReLU, LeakyReLU, Sigmoid, Softmax)
• Loss functions (MSE, CrossEntropy)
• Automatic loss function selection based on output layer
• JSON serialization/deserialization
• Comprehensive test coverage

Node.js Application

• Train neural networks on image datasets
• Recognize images with detailed statistics
• Support for custom activation functions per layer
• Automatic model saving and loading

React Application

• Interactive drawing interface
• Real-time digit recognition
• Probability visualization
• Modern, responsive UI

Quick Start

Library Usage

import { Network, LayerConfig } from "./lib/src/network.ts";
import { ActivationFunctionCollection } from "./lib/src/functions/activation.ts";

// Create network configuration
const layerConfigs: LayerConfig[] = [
	{ neurons: 784, activationFunction: ActivationFunctionCollection.ReLU },
	{ neurons: 128, activationFunction: ActivationFunctionCollection.ReLU },
	{ neurons: 10, activationFunction: ActivationFunctionCollection.Softmax },
];

// Create network (loss function auto-selected)
const network = new Network(layerConfigs);

// Train the network
network.setInputSignals(inputData).forward().backward(expectedOutput, 0.01);

Node.js Application

# Install dependencies
yarn install

# Train a new model
yarn start -t -m model.json -f ./images -l 784,128,10 -e 100 -s 0.001

# Recognize images
yarn start -r -m model.json -f ./test_images

React Application

cd example/react
yarn install
yarn dev

Architecture

The library follows a modular architecture:

• Network - Main neural network class
• Layer - Individual network layers with activation functions
• Neuron - Individual neurons with weights and biases
• Activation Functions - ReLU, LeakyReLU, Sigmoid, Softmax
• Loss Functions - MSE, CrossEntropy with automatic selection