Skip to content

TENT – adaptation on unlabeled target data by minimizing entropy #248

@davidbenjamin

Description

@davidbenjamin

Their scheme for domain adaptation, which is quite competitive and frequently used as a baseline in the literature is 1) minimize entropy on test data and 2) freeze all parameters except the affine transformation parameters of batch / layer normalization. This does not require the original source data or labels on the target data. The idea of entropy minimization is to encourage the latent representations of target and source data to align; the idea of freezing most parameters is to prevent entropy from being minimized with trivial spurious solutions.

Paper: https://arxiv.org/abs/2006.10726
This paper: https://arxiv.org/abs/2106.14999 proposes some enhancements.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions