Hackathon 3 - ML-Enhanced Analysis of EELS - Slautin

This hackathon session introduces participants to open-ended tasks in analyzing spectral image datasets, specifically electron energy loss spectroscopy (EELS) data, using machine learning techniques. The presenter outlines two notebooks: one on clustering methods and another on variational autoencoders (VAEs). The focus is on applying VAEs to EELS spectra from ITO cubes, with 12 datasets of varying density. Participants are guided through setting up a GPU runtime, loading data, and performing clustering with methods like K-means and Gaussian mixture models. The main part involves training a VAE with configurable parameters such as latent dimensionality, hidden layer sizes, learning rate, and beta (balancing reconstruction and KL divergence loss). The presenter explains the VAE architecture, including encoder, decoder, and loss functions, and emphasizes the importance of the KL divergence term for a well-structured latent space. Participants are encouraged to experiment with parameters, observe loss evolution, and visualize latent variable distributions. Two advanced tasks are proposed: testing generalization by training on one dataset and evaluating on another, and using reconstruction loss for anomaly detection. The session concludes with a brief discussion on the difference between VAEs and autoencoders, highlighting the probabilistic nature of VAEs and the potential of analyzing standard deviation in latent variables.