search

BayesWatch / deep-kernel-transfer

Official pytorch implementation of the paper "Bayesian Meta-Learning for the Few-Shot Setting via Deep Kernels" (NeurIPS 2020)
https://arxiv.org/abs/1910.05199
200 stars 29 forks source link

Why the parameters are updated on the support set data? #22

Closed 2020213484 closed 11 months ago

2020213484 commented 11 months ago

I think it will be better to update the parameters on the query set, rather than on the support set. For the reason that the evaluation process on the query set is the simulation for the meta-test setting.

mpatacchiola commented 11 months ago

Hi @2020213484, from an empirical point of view using both support and query sets for the update leads to significantly better performance. We have observed this in our experiments and it has been confirmed in a separate work.