A tiny neural network written in clojure that learns to recognize mnist digits,
without any libraries except clojure.core.matrix
.
Purely for fun and learning.
Run it with
lein run
You should see the accuracy rising
Training epoch 0
1427/2000 correct classifications
Training epoch 1
1675/2000 correct classifications
...
Training epoch 10
1728/2000 correct classifications
Test it with
lein test
Data is expected as two dimensional clojure vector. Because i was lazy i have just dumped the 10000 images from here, that is why the network currently just reaches 90% accuracy.
- load mnist from binary files
- improve accuracy with advanced techniques
- add possibility to classify an image file with the learned model