C ML

Purpose

This project is an implementation of a classical neural network in plain C, with no dependencies other than the C standard library.

Requirements

• gcc

Build Instructions

Use ./scripts/dbuild.sh and ./scripts/build/drun.sh to run the debug version and ./scripts/rbuild.sh and ./scripts/rrun.sh for the release version.

Current Status

Machine learning is working well, but may need some more optimizations to get a good neural network.

Capabilities

• Can train on IRIS in 0.02 seconds. (Run the successcounter.py script to test this yourself ONLY IF main.c is set to run iristest instead of mnisttest.)
• Can recognize handwritten digits with 75% accuracy in under 90 seconds.

Strategies Used

• Backpropagation -- a hand-derived and hand-coded implementation of partial derivatives in a neural network.
• Learning rate adjustment -- adjusts learning rate based on the error of going a single step down in the gradient.
• Repeated descent -- calculates a gradient once and goes in that direction repeatedly to avoid expensive recalculation.
• Multithreading -- can multithread loss and gradient calculations for time-efficient calculation.