A demo project to brush up my machine learning skills by writing an image recognition neural network from scratch in C.
This uses a typical 3-layer multi-layer perceptron. Training time is about 3 seconds for 80% correct, and 30s for 96% correct against the competition MNIST set of hand-written digits, which is pretty speedy for a simple C program.
- Hand-written linear algebra code in C, refined with asm inspection. Should have somewhat better cache efficiency than typical maths starter libraries for matrix multiplication.
- Hand-written CSV parser for reading the data.
- Hand-written image output for testing the data.
Note that I didn't include the test data files because they are pretty big. I used the CSV versions which you can download from https://pjreddie.com/projects/mnist-in-csv/.
