Label smoothing / Faster training / Gaussian/Laplace loss

[cnd3617]

Dear All,

We are investigating (and trying to implement) label smoothing for the classification and regression task. The goal is to add regularization elements to the model, consequently reducing the MAE.

We are also implementing a new way of training the model by pre-computing features. The goal is to obtain a trainable (i.e. faster) model for non-using-GPU computers.

Finally, we are going to work on implementing a Gaussian/Laplace loss.

Amaury, Malo, & Côme

[abursuc]

Thank you so much for the update.
Don't hesitate to reach out in case there are blocking or unclear points.
Good luck!