Spearmint integrated Bayesian Optimization for hyper parameter tuning of Auto sparse encoder embedded with Softmax Classifier for MNIST digit Classification.
- Download and install Spearmint package (instructions are on 'https://github.com/JasperSnoek/spearmint').
- Download the MNIST dataset (from 'http://yann.lecun.com/exdb/mnist/') in the same folder with the rest of Matlab files.
- Run the spearmint optimization module.
- Implementation of Classification module is in Matlab.
- STL_opt is the matlab wrapper required for spearmint package.
- config.json is the configuration file with specifications as per the spearmint instruction.
- Bayesian Optimization used is minimizing the classification error.
- L-BFGS algorithm is used to minimize the cost function for weights training in Softmax Classifier and Sparse Auto-encoder.