My take on Andrew Ng's great Machine Learning couse.
This project contains ML solutions from scratch (no external algebra / ML libraries required) in Rust. The goal is to learn about how ML works on a deep level.
The solutions in this project serve as an example, and are probably not the most correct or efficient solution to the problem. Perhaps it is more readable than other implementations, given this is written by someone who has recently started learning about machine learning.
aicourse contains solutions for lectures 1-9 of the Machine Learning course.
This includes the modules:
matrix: Matrix operationsregression: Linear, logistic and polynomial regressionnetwork: Deep feed-forward (DFF) neural network
Gradient descent is used as the optimization algorithm. The learning rate is automaticaly adjusted.
The DFF neural network uses the sigmoid activation function and can be trained in parallel (regularization parallelism). There is no proper splitting up of train / cross-validation / test datasets, so the model optimization is not optimal.
An executable that trains a DFF neural network to classify digits based on the MNIST dataset.
To see a working neural network in action, clone the project including
submodules and run:
cargo run -p aicourse-train --release
To see the available options, run:
cargo run -p aicourse-train --release -- --help
After training a [28 * 28, 256, 10] unit network in parallel for ~70 epochs with the first 5000 samples of the MNIST train dataset, it is able to classify the MNIST test dataset with an accuracy of ~92%. This takes approximately 7 minutes with a Ryzen 5 1600 CPU.
A program that reads a video feed from a webcam and classifies digits in
real-time. Uses a neural network configuration from aicourse-train.
Currently only Linux is supported. To run the program:
cargo run -p aicourse-webcam --release