Live training & inference playground β see weights, activations, and outputs in real time.
Idle
Inputs β Interactive Sliders
Training Weights: Not Transferred
Move sliders β inference runs instantly and records to history. Loading a weights CSV updates architecture & inference model.
Output Plots β Current vs History
Hover to see values. Lines show each output zβ across inferences; bars show the latest inference only.
Output History β Tabulated
Training Data β Tabulated
Shows generated or CSV-loaded dataset. For large datasets, increase βRows to showβ.
Training β Dataset, Hyperparameters & Learn
Loss is MSE. Transfer weights to apply them in the Inputs panel.
Neural Net β Weights & Activations
Blue=positive, red=negative. Pulse = recent inference. Hover nodes for details.
Weights β Tabulated
Shows weights loaded from CSV or currently trained (via Save Config + Weights).
Help β Visual MLP Studio
What is this?
This app demonstrates a one-hidden-layer Multi-Layer Perceptron (MLP) with live training and inference. You can generate or load training data, train, visualize weights/activations, and run inference via sliders.
Panels
Inputs: colored sliders for each input xα΅’. Moving them runs inference instantly and records history. Use Load Weights CSV to import previously exported mlp_weights.csv.
Output Plots: bar chart for the latest inference or multi-line chart for history. Hover for exact values.
Output History: a table of all inferences (x and z). βExport CSVβ downloads the table.
Training Data: shows the current dataset (generated or CSV). You can limit visible rows and export CSV.
Training: architecture, activation, hyperparameters, data generation/loading, training controls, and loss curve.
Neural Net: graph of weights; line thickness and color encode weight magnitude/sign. A pulse runs each inference.
Weights β Tabulated: raw weight/bias entries (type, from, to, value). Useful for auditing & diffing weight sets.
Training Settings
Architecture (X / Y / Z): number of inputs, hidden units, and outputs. Changing it resets weights and history.
Activation (hidden): tanh, relu, leakyrelu, sigmoid, elu, gelu (output layer is linear).
Learning Rate: gradient step size.
Epochs per Click: passes over the dataset per train click.
Batch Size: samples per update.
Dataset Size: size for generated datasets.
Generate Dataset: Random Linear Map or XOR (X=2, Z=1).
Load CSV: expects rows with X input columns followed by Z target columns.