读书笔记

FEDERATED LEARNING FROM ONLY UNLABELED
DATA WITH CLASS-CONDITIONAL-SHARING CLIENTS
1.problems:
2.

REPRESENTATIONAL CONTINUITY FOR
UNSUPERVISED CONTINUAL LEARNING
1.problems:
(1) Although Rao et al. (2019) instantiated a continual
unsupervised representation learning framework (CURL), it is not scalable for high-resolution tasks,
as it is composed of MLP encoders/decoders and a simple MoG generative replay. This is evident in
their limited empirical evaluation using digit-based gray-scale datasets.
(2) However, a common assumption for
existing methods is the availability of a large amount of unbiased and unlabelled datasets to learn
the feature representations.

Barely-Supervised Learning:
Semi-supervised Learning with Very Few Labeled Images
1.problem: