← Main Site Neural Lab

Paused

NEURAL-LAB

Neural Lab

Learn deep learning through interactive architecture, mechanism, and training visualizations.

Live Architecture Preview 3D structure · 2D diagram · signal flow

Active Layer Hidden 02

Signal Flow Input → Latent → Output

Loss Curve

Update Step θ ← θ - η∇L

Suggested Curriculum

Follow the model from structure to learning

Six focused modules move from network anatomy to the math of one update. Each stage opens the matching lab and gives you a concrete thing to inspect before moving on.

6 stages Architecture -> Mechanisms -> Training

Simulation

Step 0

Epoch 0

Network Type

Speed: 1

Architecture Parameters

MLP Parameters

Performance

Particles Scene Labels Glow: 100% Animation Density: 100% Connection Lines: 100%

Architecture MLP

Camera Controls

Zoom

Last action: Ready

Guidance

Auto Info Cards Auto Architecture Spotlight

Legend

Input / source data

Hidden processing

Output / decision

Latent / gates / memory

Moving particles show signal flow

Lines show weighted connections

Training Cycle

Training Cycle Controls

Speed

Lesson Controls

Dataset

x₁ 0.700 x₂ -0.400 Target y 1.000

A clear positive example with target 1.00.

Training Mode

Single Update Preview keeps the beginner lesson focused: every final step shows what would change without changing the actual weights.

Learning Rate Learning Rate η 0.300

0.300 creates a moderate final-step update and lowers this sample's loss after one update.

Preview Update Compare Update Show prediction, loss, and key parameters before vs after one step.

MLP Training Tutorial

End-to-End Learning Cycle

Step 01: Data Enters

One labeled example enters the network. The feature vector becomes the visible starting signal, while the target waits at the end as the answer the model should move toward.

Math Focus

Formula

Why It Matters

This stage is only setting context: the parameters are fixed, and no activation, loss, or gradient has been produced yet.

What Changes

Before This Step

The network is idle with fixed starting weights and biases.

After This Step

One labeled sample is loaded, so the input vector and target are now the active training context.

The system moves from a blank lesson state to a concrete example the model can process.

NEURAL-LAB

NEURAL-LAB

NEURAL-LAB

Neural Lab

Follow the model from structure to learning

Simulation

Network Type

Architecture Parameters

Performance

Camera Controls

Guidance

Legend

ReLU

Loss Surface Descent

Mean Squared Error

Diffusion

End-to-End Learning Cycle

Welcome to Neural Lab