Learning to reject unknown samples (not present in the source classes) in the
target domain is fairly important for unsupervised domain adaptation (UDA).
There exist two typical UDA scenarios, i.e., open-set, and open-partial-set,
and the latter assumes that not all source classes appear in the target domain.
However, most prior methods are designed for one UDA scenario and always
perform badly on the other UDA scenario. Moreover, they also require the
labeled source data during adaptation, limiting their usability in data
privacy-sensitive applications. To address these issues, this paper proposes a
Universal Model ADaptation (UMAD) framework which handles both UDA scenarios
without access to the source data nor prior knowledge about the category shift
between domains. Specifically, we aim to learn a source model with an elegantly
designed two-head classifier and provide it to the target domain. During
adaptation, we develop an informative consistency score to help distinguish
unknown samples from known samples. To achieve bilateral adaptation in the
target domain, we further maximize localized mutual information to align known
samples with the source classifier and employ an entropic loss to push unknown
samples far away from the source classification boundary, respectively.
Experiments on open-set and open-partial-set UDA scenarios demonstrate that
UMAD, as a unified approach without access to source data, exhibits comparable,
if not superior, performance to state-of-the-art data-dependent methods.

该文提出了一种名为 UMAD 的通用模型适应框架，能够在不使用源数据或了解域之间类别偏移的先验知识的情况下处理 open-set 和 open-partial-set 的两种 UDA 场景，并通过实验证明其在数据隐私敏感型应用中具有可比较甚至更优的性能。