Neural Networks From Scratch

A neural network built from scratch using only NumPy — no PyTorch, no TensorFlow, no autograd. Based on the Neural Networks from Scratch book by Harrison Kinsley, extended with an interactive Flask web GUI for live visualization.

What's in here

Core network — neuron_test.py

Full forward pass pipeline built by hand:

• Dense (fully connected) layers
• ReLU activation
• Softmax activation with numerical stability (subtracts row max before exp to prevent overflow)
• Categorical cross-entropy loss — handles both sparse integer labels and one-hot encoded labels

Interactive GUI — neural_network_gui.py

Flask web app wrapping the network with 4 live visualizations:

• Original spiral data
• Network predictions vs ground truth
• Per-class prediction confidence heatmap
• Per-sample loss curve

Custom data generator — spiral_data.py

Hand-written n-class spiral dataset generator with configurable noise.

Setup

pip install numpy nnfs flask matplotlib

Usage

Run the core network:

python neuron_test.py

Launch the web GUI:

python neural_network_gui.py

Open http://localhost:5000, set the number of samples and classes, and hit Train.

Implementation breakdown

Component	Details
Layer_Dense	Weight init 0.01 * randn, bias init zeros, forward via dot product
Activation_ReLU	Elementwise max(0, x)
Activation_Softmax	Numerically stable — row-max subtraction before exp
Loss_CategoricalCrossentropy	Clipped predictions, handles sparse and one-hot targets
Spiral generator	Parameterized by points and classes, Gaussian noise on angle

Stack

Python · NumPy · Flask · Matplotlib