Multiclass classification

[chris-nemeth]

We can currently handle binary classification through GPMC and the BernLik likelihood. It would be good to extend this functionality to the multiclass setting.

[maximerischard]

We've already tagged v0.6.0. Should this be for v0.7.0?

[chris-nemeth]

Upated

[maximerischard]

So how would this work? Do we need more general support for multivariate GPs?

[chris-nemeth]

I need to revisit this again as it may be as hard to implement as I first thought. My initial concern was that we'd have to change our latent variables v in GPMC from a vector to a matrix, where we'd introduce a new column for each possible class. However, it might be possible to avoid this change by vectorising the latent variables and using a block diagonal covariance matrix, where each block corresponds to each class, i.e. 3 blocks if you have a 3-way multiclass.