Latent features in analysis

[pdeblanc]

This is a feature request. I'd like an option to return latent feature activation maps in the JSON returned by the analysis engine. This only applies to the evaluation of the root position, not any of the search tree positions.

There are three applications I have in mind for this:

1. Interpretability experiments: by visualizing activations in a tool like KaTrain, we may gain insight into how the neural network is reasoning about positions.
2. Neural pattern search: feature vectors could be used to find similar positions (e.g. measured by L2 norm) in a corpus of games. This would be similar to existing joseki and fuseki search tools like Waltheri, but without requiring an exact match. This might allow us to find instances of a given tesuji or life-and-death shape in games, which seems inherently fuzzier than joseki search.
3. Problem categorization: clustering on feature vectors could be used to organize a collection of tsumego by theme.

[hadim]

That would be great to be able to extract a latent vector from a go board indeed.