Distribution Alignment Using Complement Entropy Objective and Adaptive Consensus-Based Label Refinement For Partial Domain Adaptation

In this work, we address a realistic case of unsupervised domain adaptation, where the source label set subsumes that of the target. This relaxation in the requirement of an identical label set assumption, as witnessed in the standard closed-set variant, poses a challenging obstacle of negative transfer that potentially misleads the learning process from the intended target classification objective. To counteract this issue, we propose a novel framework for a partial domain adaptation setup that enforces domain and category-level alignments through optimization of intra- and inter-class distances, uncertainty suppression on classifier predictions, and target supervision with an adaptive consensus-based sample filtering. In this work, we aim to modify the latent space arrangement where samples from identical classes are forced to reside in close proximity while that from distinct classes are well separated in a domain-agnostic fashion. In addition, the proposed model addresses a challenging issue of uncertainty propagation by employing a complement entropy objective that requires the incorrect classes to have uniformly distributed low-prediction probabilities. Target supervision is ensured by employing a robust technique for adaptive pseudo-label generation using a nonparametric classifier. The methodology employs a strategy that permits supervision from target samples with prediction probabilities higher than an adaptive threshold. We conduct experiments involving a range of partial domain adaptation tasks on two benchmark datasets to thoroughly assess the proposed model’s performance against the state-of-the-art methods. In addition, we performed an ablation study to validate the necessity of the incorporated modules and highlight their contribution to the proposed framework. The experimental findings obtained manifest the superior performance of the proposed model when compared to the benchmarks.