Image Reduction based on Machine Learning

Our goal was to compress labeled image data using MATLAB's built-in autoencoders (a data-reduction program) to prioritize each image's label.

• The typical autoencoder uses unlabeled images/data, reducing the data to have the least amount of error when reconstructing the original data.

• Our goal was to use the standard autoencoder with labeled images, and reduce the data to have the least amount of error when classifying the data.

• Our solution was to append the label to our autoencoder training data, where the label would in essence become a part of the reconstruction error that it was trying to optimize.

• For our testing data, which would have to be unlabeled by nature, we would simply append a label of zero to our data so that the autoencoder could still run on it.

• Our results shows some promise and that this idea could be optimized to work better than an unmodified autoencoder, but generally our results were inconsistent and limited.

• Our GitHub repository contains the our final report, which has specifics of our reasoning and results.