[Paper Review] Semi-supervised Domain Adaptation via Minimax Entropy

Kuniaki Saito, Donghyun Kim+3 more|arXiv (Cornell University)|Apr 13, 2019

Domain Adaptation and Few-Shot Learning39 references76 citations

TL;DR

Introduces Minimax Entropy (MME), an adversarial, entropy-based SSDA method that learns domain-invariant prototypes and discriminative features by alternating entropy maximization on unlabeled target data with entropy minimization in the feature extractor.

ABSTRACT

Contemporary domain adaptation methods are very effective at aligning feature distributions of source and target domains without any target supervision. However, we show that these techniques perform poorly when even a few labeled examples are available in the target. To address this semi-supervised domain adaptation (SSDA) setting, we propose a novel Minimax Entropy (MME) approach that adversarially optimizes an adaptive few-shot model. Our base model consists of a feature encoding network, followed by a classification layer that computes the features' similarity to estimated prototypes (representatives of each class). Adaptation is achieved by alternately maximizing the conditional entropy of unlabeled target data with respect to the classifier and minimizing it with respect to the feature encoder. We empirically demonstrate the superiority of our method over many baselines, including conventional feature alignment and few-shot methods, setting a new state of the art for SSDA.

Motivation & Objective

• Demonstrate limitations of unsupervised/domain alignment methods in SSDA settings with few target labels.
• Propose a minimax entropy framework to adapt classifiers by leveraging unlabeled target data.
• Develop a cosine-similarity prototype-based classifier whose prototypes are adapted via entropy-based objectives.
• Show that MM E outperforms baselines on DomainNet, Office-Home, and Office datasets across one-shot and three-shot settings.

Proposed method

• Adopts a similarity-based network with a feature extractor F and a cosine-similarity classifier C using class prototypes W.
• Estimates domain-invariant class prototypes by maximizing the entropy of predictions on unlabeled target data (prototype adaptation).
• Minimizes entropy with respect to the feature extractor to cluster unlabeled target features around prototypes (discriminative clustering).
• Trains F and C with a standard classification loss on labeled source and target data, plus a minimax entropy term controlled by a hyper-parameter lambda.
• Implements the minimax game via a gradient reversal layer to jointly optimize entropy maximization (for C) and entropy minimization (for F).
• Evaluates on DomainNet, Office-Home, and Office datasets in one-shot and three-shot target-label regimes, reporting improvements over baselines.

Experimental results

Research questions

• RQ1 Do existing UDA/SSL methods improve SSDA performance when only a few labeled target examples are available?
• RQ2 Can an adversarial, entropy-based approach to estimating domain-invariant prototypes improve target-domain discrimination?
• RQ3 Does maximizing entropy on unlabeled target data to adjust prototypes, followed by entropy minimization in the feature extractor, reduce domain divergence and improve accuracy?
• RQ4 How does MM E compare to standard entropy minimization and domain-alignment baselines across multiple datasets and network backbones?

Key findings

• MME achieves state-of-the-art performance for SSDA across multiple benchmarks and architectures.
• Compared to baselines, MME reduces error relative to simple source-target training by substantial margins in one-shot and three-shot settings.
• Entropy-based prototype adaptation combined with featureExtractor clustering yields improved domain alignment and discriminative target features.
• MME consistently outperforms unsupervised domain adaptation methods that rely solely on feature distribution alignment in SSDA scenarios.
• Visualization and analysis show reduced domain divergence and more compact, prototype-centered target feature clusters under MME.