Skip to content

A version of K-Nearest Neighbors that uses cosine similarity as a distance metric, and classifies MNIST data at almost twice the rate of standard K-NN implementations

License

Notifications You must be signed in to change notification settings

samueleverett01/K-Nearest-Neighbors-Image-Classifier

Repository files navigation

Image Classifier

An implementation of a parsimonious K-NN model that yeilds a sizeable performance boost over standard K-NN algorithms from Scikit-Learn.

I placed all the notebooks I used to create this model into the "Model Development" folder for my reference.
The final model, along with its analysis and comparison to the K-NN model offered by Scikit-Learn is in the "Cosine-Similarity Model Analysis" notebook.

* Updates

The Randomized K-NN and Superior K-NN notebooks seek to implement a K-NN style algorithms that leverage methods analogous to iDistance for image classification. These two notebooks show that these new classification algorithms classify datapoints far faster than standard "brute force" and tree based methods (like KD-Tree or Ball Tree), however they unfortunately have relatively low classification accuracy.

About

A version of K-Nearest Neighbors that uses cosine similarity as a distance metric, and classifies MNIST data at almost twice the rate of standard K-NN implementations

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published