An AI trained on the mnist dataset. The code is broken, but it does kinda work. In the first step of backpropagation, they AI will greatly reduce the error. But after that, the AI will slowly get worse at guessing. You are tho able to generate images, and some networks like the network_100.json
shows the patterns it's looking for when generating those images, even tho its a very bad network, with only about 10% accuracy. The best network right now is the network_low_lr.json
which has about 72% accuraccy on the test dataset.
You can use the --help
flag when running the binary to see available features and usage. Make sure to have the images
, config
, data
and logs
folders, in the same folder as the binary.
cargo run -- -m test -i network_low_lr.json --generate-images
cargo run -- -m train -o network.json --iterations=1000
cargo run -- -m train -o network.json -l 0.01