Road Object AI Training Simulator

Training controls

Speed 2×

Speed tip. At higher speeds the simulator shortens pauses and also runs extra fast training cycles between animations, so visible learning outcomes appear sooner.

The data

18 training examplesUSED TO LEARN

The network updates its weights from labelled road-picture clues.

9 unseen checksNOT USED TO LEARN

These are only used to measure unseen accuracy, never to update weights.

CAR

Car

CYC

Cyclist

SIGN

Road Sign

Training progress

Epoch0all 18 training examples once

Model size95Layer 1: 6 · Layer 2: 5

Training accuracy33%examples used for learning

Unseen accuracy33%separate check examples

Model size = trainable numbers. This version uses two hidden layers. This simulator now uses a fixed architecture during training: the hidden layers keep the same number of neurons while the weights and biases are updated. Real camera systems would usually use many more input values, deeper models and far larger training datasets.

Connection key

Blue = positive learned weight.

Coral = negative learned weight.

Thicker = larger weight magnitude.

Road Camera AI · Training Loop One spotlight example is animated; the remaining examples in the epoch use the same update rule off-camera

TRAINING

Spotlight objectCar

Correct objectCar

Network predictionWaiting

Current stageReady

HIDDEN LAYERS

the network learns useful patterns here

INPUTS

HIDDEN LAYER 1

HIDDEN LAYER 2

OUTPUTS

Wheel cues

round features

Reflections

glass / shine

Body shape

vehicle outline

Human form

rider / person

Bar / pole

handlebar cue

Sign edges

frame geometry

Car33%

Cyclist33%

Road Sign33%

Ready for real training.

The network starts small with two hidden layers. Each epoch uses all 18 training examples.

Correct-class probabilityBefore: —

After this updateAfter: —

What is happening?

Ready to train

The model begins with a fixed architecture and random weights. Training repeatedly predicts the road object, measures the error and updates the weights and biases.

Ordinary training changes weights and biases. In this version, the architecture stays fixed throughout training.

The real training loop

Input: one labelled road-picture example enters.

Forward pass: the network calculates probabilities for the road objects.

Loss: measure how far the prediction is from the correct label.

Backpropagation: send information about that error backwards.

Optimiser: make a small change to the weights and biases.

Architecture choice

Architecture is fixed during training. In a normal neural network, you choose the number of hidden layers and neurons before training starts. Training then improves the weights and biases inside that fixed structure.

The important accuracy

Training accuracy can rise because the model has seen those examples. Unseen accuracy is the safer check because those examples are never used to update weights.

Big idea

Real neural-network training means optimising weights and biases to reduce a loss. This simulator uses a simplified road-camera example, while real systems would train on much larger datasets with many more input features and a more complex architecture.