Neural Network Lab

Dataset-Driven Multi-Layer Neural Network Trainer

Train a fully-connected multi-layer neural network directly from complete input-output patterns. Configure hidden layers, activations (ReLU, LeakyReLU, Sigmoid, tanh, Linear, Softmax), initialization schemes, mini-batch SGD, adaptive learning-rate on validation plateau, and live SVG visualization of weights, biases and activations.

Simulator

Input Neurons (UI hint)

Output Neurons (UI hint)

Hidden Neurons (comma separated)

Hidden Activation

Output Activation

Learning Rate

Stop Loss Value

Training Steps

Log Interval (steps)

Evaluation Percent (0-100)

Batch Size (1 = online)

Patience (plateau steps)

LR Multiplier (on plateau)

Use -1 to 1 initialization

Use 0 to 1 initialization

Use 0 or 1 initialization

Use Xavier initialization

Use MinMax initialization

Train Biases

Shuffle dataset before splitting

Use Cross-Entropy loss for softmax

Complete Patterns (newline separated — inputs + outputs)

Network Visualization

Training Loss

Training Log

Ready. Paste patterns and click “Begin Simulation”.

Machine-Readable NN Definition

Evaluation Results

Acknowledgments

Special thanks to Grok for its invaluable assistance in creating this neural network simulator for the Deep Inside workshop series.