Adversarial Discriminative Domain Adaptation

TL;DR

Addresses unsupervised domain adaptation under domain shift by learning a separate target encoder that matches the source feature distribution through adversarial learning. The authors present ADDA, an asymmetric, unshared-weight instantiation using a GAN loss within a unified framework that subsumes prior methods. Empirical results on MNIST, USPS, SVHN, and RGB-depth NYU show competitive or superior performance, including cases where GAN-based domain transfer outperforms generative approaches. The work suggests that discriminative adaptation with asymmetric mappings and GAN losses can achieve strong cross-domain generalization without generating target-domain images.

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.

Figures (5)

• Figure 1: We propose an improved unsupervised domain adaptation method that combines adversarial learning with discriminative feature learning. Specifically, we learn a discriminative mapping of target images to the source feature space (target encoder) by fooling a domain discriminator that tries to distinguish the encoded target images from source examples.
• Figure 2: Our generalized architecture for adversarial domain adaptation. Existing adversarial adaptation methods can be viewed as instantiations of our framework with different choices regarding their properties.
• Figure 3: An overview of our proposed Adversarial Discriminative Domain Adaptation (ADDA) approach. We first pre-train a source encoder CNN using labeled source image examples. Next, we perform adversarial adaptation by learning a target encoder CNN such that a discriminator that sees encoded source and target examples cannot reliably predict their domain label. During testing, target images are mapped with the target encoder to the shared feature space and classified by the source classifier. Dashed lines indicate fixed network parameters.
• Figure 4: We evaluate ADDA on unsupervised adaptation across four domain shifts in two different settings. The first setting is adaptation between the MNIST, USPS, and SVHN datasets (left). The second setting is a challenging cross-modality adaptation task between RGB and depth modalities from the NYU depth dataset (right).
• Figure 5: Confusion matrices for source only, ADDA, and oracle supervised target models on the NYUD RGB to depth adaptation experiment. We observe that our unsupervised adaptation algorithm results in a space more conducive to recognition of the most prevalent class of chair.