Embeddings Analysis

[daniel-z-kaplan]

PCA analysis is a common method of looking at embeddings spaces. It's a potentially relevant method here for our visualization.

For this task, we would like to take data from existing benchmarks (such as BACH), and embedd it, using a current checkpoint from the Midnight replication (this will be provided). These embeddings are also available after the evaluation process, natively.

We will try at least two different methods:

• PCA analysis using top 4 components (matching the number of classes),
• Clustering + coloration

[sagarigrandhi]

Interested!