How to Build a Breast Cancer Prediction Neural Network from Scratch in Python

In constructing a breast cancer prediction neural network, the process is divided into three major parts: (1) creating a neural network from scratch in Python and training it with gradient descent; (2) applying the Wisconsin breast cancer dataset with nine features to classify tumors as benign or malignant; (3) exploring how back‑propagation and gradient descent work.

While many powerful libraries like TensorFlow, PyTorch, Fast.ai, Keras, MXNet, DL4J, etc., exist, building the model manually helps uncover essential concepts that would be missed by using high‑level APIs alone.

Understanding Functions and Supervised Learning

Functions map inputs to outputs, and in machine learning three common learning types exist: supervised learning (learning from labeled data), unsupervised learning (learning from unlabeled data), and reinforcement learning (learning from rewards). This tutorial focuses on supervised learning, using a dataset of input‑output pairs to discover the underlying function.

Simple Two‑Layer Neural Network

A basic network consists of an input layer (matching the number of features, e.g., nine for the Wisconsin dataset), a hidden layer, and an output layer. Although deeper networks are possible, a two‑layer structure is sufficient to illustrate core ideas.

Each neuron has a weight and a bias. The linear combination is computed as W*X+b, where W represents weights, X the input vector, and b the bias term.

Activation Functions

Linear functions alone cannot model complex, non‑linear relationships, so activation functions are added after each linear transformation. Four typical activation functions are discussed:

Sigmoid : 1/(1+e**-x), output range [0,1], suitable for binary classification but prone to gradient vanishing.

Tanh : (2/(1+e**-2x))-1, output range [-1,1], steeper than sigmoid but shares similar drawbacks.

ReLU : max(0,x), output range [0,∞), simple and efficient, yet can cause dead neurons when inputs are zero.

Leaky ReLU : normalizes inputs into a probability distribution, often used in multi‑class output layers.

These functions introduce non‑linearity, enabling the network to approximate complex mappings, but they also bring challenges such as gradient vanishing and exploding.

Putting It All Together

For a two‑layer network, the forward computation becomes:

Ŷ = A2 = Sigmoid(W2*ReLU(W1*X + b1) + b2)

The network must learn the unknown parameters W and b. Training involves initializing these parameters randomly and then iteratively updating them using gradient descent based on the loss between the predicted output Ŷ and the true target Y.

Finally, the tutorial hints at implementing the entire process in Python by defining a class that handles parameter initialization and the forward‑backward passes.