How mrassell/homemade-neural-network works
Project type and purpose
• This is an educational, from-scratch neural-network implementation in a single Python script (NeuralNetwork.py) that trains and evaluates a two-layer neural network on the MNIST dataset.
• There are no separate frontend/backend or microservices; it’s a standalone data-processing and ML pipeline.
File structure
• Top-level file: NeuralNetwork.py
• No subdirectories—everything lives in one script.
• External dependencies are defined in the README: NumPy, pandas, matplotlib.
README for architecture clues
• Describes stages: data loading/preprocessing, parameter initialization, forward propagation, backward propagation (gradient computation), parameter updates (gradient descent), and testing/visualization.
• Key functions map directly to pipeline steps.
Main components
a. Data Layer
• MNIST CSV file (train.csv)
• pandas for data loading
• preprocessing to normalize pixel values and split into train/dev sets
b. Model Component
• Parameter storage: W1, b1, W2, b2
• Activation functions: ReLU (hidden), Softmax (output)
• Forward propagation module
• Backward propagation module
c. Training Engine
• Loss calculation (cross-entropy)
• Gradient descent loop (iterations + learning rate α)
• Logging training accuracy every N steps
d. Evaluation & Visualization
• Test predictions on sample images
• matplotlib for displaying images and predicted labels
e. External Libraries
• NumPy (numerical computation)
• pandas (data I/O)
• matplotlib (plotting)
Relationships and interactions
• From Data Loader to Preprocessor
• Preprocessed data flows into Training Engine
• Training Engine repeatedly invokes Model’s Forward → Loss → Backward → Parameter Update
• After training, final parameters go into Evaluation module
• Evaluation module outputs images + predicted labels via matplotlib
Architectural patterns/design principles
• Pipeline pattern: data → model → training loop → evaluation
• Clear separation of concerns even in one file: data I/O vs. core model vs. training logic vs. visualization