[Paper Review] Adversarial Discriminative Domain Adaptation

Eric Tzeng, Judy Hoffman+2 more|arXiv (Cornell University)|Feb 17, 2017

Domain Adaptation and Few-Shot LearningComputer Science16 references454 citations

TL;DR

ADDA introduces an unsupervised domain adaptation method with discriminative source representations, unshared (asymmetric) target mappings, and a GAN-based adversarial loss, achieving state-of-the-art results on cross-domain digit datasets and cross-modality RGB-to-depth tasks.

ABSTRACT

Adversarial learning methods are a promising approach to training robust deep networks, and can generate complex samples across diverse domains. They also can improve recognition despite the presence of domain shift or dataset bias: several adversarial approaches to unsupervised domain adaptation have recently been introduced, which reduce the difference between the training and test domain distributions and thus improve generalization performance. Prior generative approaches show compelling visualizations, but are not optimal on discriminative tasks and can be limited to smaller shifts. Prior discriminative approaches could handle larger domain shifts, but imposed tied weights on the model and did not exploit a GAN-based loss. We first outline a novel generalized framework for adversarial adaptation, which subsumes recent state-of-the-art approaches as special cases, and we use this generalized view to better relate the prior approaches. We propose a previously unexplored instance of our general framework which combines discriminative modeling, untied weight sharing, and a GAN loss, which we call Adversarial Discriminative Domain Adaptation (ADDA). We show that ADDA is more effective yet considerably simpler than competing domain-adversarial methods, and demonstrate the promise of our approach by exceeding state-of-the-art unsupervised adaptation results on standard cross-domain digit classification tasks and a new more difficult cross-modality object classification task.

Motivation & Objective

• Motivate the need for robust unsupervised domain adaptation under domain shift and dataset bias.
• Propose a unified framework for adversarial domain adaptation and identify design choices.
• Introduce ADDA as a discriminative, unshared-weight, GAN-based instantiation.
• Demonstrate ADDA’s effectiveness on cross-domain digit datasets and cross-modality RGB-to-depth tasks.

Proposed method

• Present a generalized framework for adversarial unsupervised adaptation.
• Use a domain discriminator D to distinguish source vs target encoded features.
• Employ an asymmetric target encoder M_t with unshared weights from the source encoder M_s.
• Apply a GAN-style adversarial loss to train M_t to fool D (inverted labels GAN loss).
• Initialize M_s and classifier C_s with labeled source data and fix M_s during target adaptation.
• Demonstrate that discriminative modeling suffices without image-space generators.

Experimental results

Research questions

• RQ1 Can a discriminative, asymmetric, GAN-based adaptation framework achieve competitive unsupervised domain adaptation without image generation?
• RQ2 Does untying weights between source and target encoders improve adaptation over symmetric shared-weight approaches?
• RQ3 How does ADDA perform on standard cross-domain digit shifts and cross-modality RGB-to-depth shifts?
• RQ4 What are the effects of different adversarial losses within a unified framework for domain adaptation?

Key findings

• ADDA achieves state-of-the-art or competitive results on MNIST, USPS, and SVHN unsupervised adaptation in three-direction digit tasks.
• In MNIST→USPS, USPS→MNIST, and SVHN→MNIST, ADDA attains 0.894±0.002, 0.901±0.008, and 0.760±0.018 respectively.
• On NYU depth RGB-to-depth cross-modality adaptation, average accuracy improves from 13.9% (source only) to 21.1% after ADDA.
• ADDA demonstrates strong cross-domain and cross-modality transfer without requiring labeled target data or image-space generators.
• The method outperforms several competing adversarial approaches in multiple domain shifts and provides insights into the benefits of asymmetric mappings.