NSF Research Progress Summary

Complex Systems Project 2 (Wall Paintings)

Sponsor Organization: the National Science Foundation

Supervising Faculty: Dr. Luis Rocha

Collaborators: Yingnan Ju

We tried different methods of unsupervised machine learning for classification of wall paintings: Traditional K-means and Autoencoding + K-means. Traditional K-means method uses all raw pixels of each image as the input and divide a batch of images into n (here n=4) categories and each category contains images that are most similar with each other.

Another method is to combine Autoencoding and K-means together. Neural network is applied as an autoencoder and it compresses an image into k (like, k=20) numbers and with only these k numbers, the neural network can also decode them to restore or reconstruct a nearly same original image. With this method, the number of features could be reduced to 20 features from 30000 features. Although there is only 20 features, these features might be the most useful for computer because the neuron network could reconstruct the image with only these 20 numbers.

Image from: https://towardsdatascience.com/applied-deep-learning-part-3-autoencoders-1c083af4d798  

In simple words, the neuron network thinks these 20 features are the most important and useful. Then I can use K-means to deal with these 20 features to get the result with using these 20 features as the input of K-means to classify these wall paintings.

\\\\