Self influence

[tfjgeorge]

No description provided.

[tfjgeorge]

in summary, this unifies the @linear api to a single LinearModel class, with 'log_loss' and 'l2' loss currently implemented:

• adds tests vs LOO
• adds support for binary classification (whereas previously it was handled as a multiclass classification using softmax)
• grad_y and grad_p methods
• the influence probe is renamed SelfInfluence
• removes untested classifiers
• updates TracIn and GradSimilarity in accordance

Possible regression: sparse feature matrices

[pierrenodet]

Overall looks very clean :)

[pierrenodet]

It looks good (besides trees, but we can look into that later), but what's up with the scales in the last plot?

[tfjgeorge]

no idea :-D I agree that it looks suspicious