Neural Network Lab

Neural Network Pattern Lab

A Neural Network recognises cancer from tumour measurements the way a radiologist recognises patterns from a scan — by extracting increasingly abstract features through layers of neurons.

Enter patient data and press ▶ Forward Pass to watch signals flow through the network.

Forward Pass & Activation

Each neuron computes a weighted sum of its inputs, then applies an activation function to introduce non-linearity:

z = Σ (wᵢ · xᵢ) + b [weighted sum] ReLU(z) = max(0, z) [hidden layers] σ(z) = 1/(1+e⁻ᶻ) [output layer] Output = probability of malignancy

Backpropagation & Learning

After a forward pass, if the prediction is wrong, the error is propagated backwards to adjust every weight via gradient descent:

Loss = −[y·log(ŷ) + (1−y)·log(1−ŷ)] (Binary Cross-Entropy) ∂Loss/∂w = chain rule of calculus w ← w − η · ∂Loss/∂w (gradient descent update)

⚖️ The Black Box Problem

A neural network with 3 hidden layers and 512 parameters cannot explain why it flagged a patient as malignant. A clinician cannot use "the third hidden neuron fired strongly" as a medical justification. This is the core challenge of clinical AI explainability.

① Patient Input

Tumour Size

centimetres (cm)

Cell Uniformity

score 1 – 10

Hidden Layers2

neurons per layer: 4

Learning Rate (η)0.01

step size for weight updates

② Network Architecture

Layer computation:
z = Σ wᵢ·xᵢ + b
a = ReLU(z) = max(0, z)
Output: σ(z) = 1/(1+e⁻ᶻ)

Input: size, uniformity
Hidden: 2 × 4 neurons (ReLU)
Output: 1 neuron (Sigmoid)
Parameters: —

③ Prediction Output

Malignancy Probability

MALIGNANT

—

—

Output z-score

—

BCE Loss

—

Layer 1 Max Act

—

Layer 2 Max Act

④ Network Visualization — node brightness = activation level · edge thickness = weight magnitude · animated signal flow on Forward Pass