Projects · University
Neural Network from scratch
A multithreaded feed-forward neural network in C++26 with no ML libraries, trained on UCI Car Evaluation and MNIST. Submitted as my algorithms and data structures final project.
$ ./build/NeuralNetwork
Available data:
- car.data
- mnist
- Randomly Generated Data
Forward pass, backpropagation and training loop written by hand, parallelised across threads, with a test suite.
- Car Evaluation (UCI): 1,728 samples, six categorical attributes one-hot encoded into 21 inputs, four classes. Trains to about 100% test accuracy quickly.
- MNIST handwritten digits from CSV.
- Synthetic data whose target is the index of the largest input, useful for smoke-testing a topology without anything on disk.
Trained weights are written to disk as binary files.
