/MNIST-challenge

The MNIST dataset tackled by several machine learning models coded from scratch

Primary LanguagePythonMIT LicenseMIT

Description

The MNIST dataset is targeted by several machine learning models written from scratch. The implemented algorithms so far are: vanilla kNN, logistic regression, convolutional neural networks and classification Gaussian processes. The code for the neural networks is partly influenced by the Stanford CS231 assignments.

Requirements

It is tested in Anaconda 2.7 environment. Mainly the external packages numpy, scipy and matplotlib are used, so an Anaconda installation is not really required.

Run

see main.py

Results:

accuracies on the test set:

  • kNN: 97.05 %
  • LogReg: 92.61 %
  • ConvNet: 99.03 %
  • CGPs: 95.39 %