Abstract
Domain adaptation (DA) aims to transfer knowledge from a label-rich source domain to a label-scarce target domain. While CycleGAN-based methods have shown promise by enforcing pixel-level cycle-consistency, they often struggle in low-resource scenarios where rigid reconstruction constraints can lead to instability and overfitting to domain-specific noise. This paper introduces Augmented Cyclic Adversarial Learning (ACAL), which replaces traditional pixel-level reconstruction with a task-specific semantic consistency. By relaxing the cycle-consistency constraint and augmenting the discriminator with task-specific feedback, ACAL achieves significant performance gains in both visual and auditory domain adaptation tasks, particularly when target data is extremely limited.
Introduction
The challenge of domain adaptation lies in bridging the distribution gap between source and target domains. Standard approaches often utilize CycleGANs to learn bidirectional mappings. However, the authors identify a critical bottleneck: the reliance on $L_1$ or $L_2$ pixel-level reconstruction losses. In low-resource settings, forcing a model to reconstruct every pixel exactly is often counterproductive, as the model may prioritize matching irrelevant style details over preserving essential task-specific features. This "reconstruction-is-too-hard" problem leads to poor generalization and unstable training.
Methodology: Augmented Cyclic Adversarial Learning (ACAL)
The core innovation of ACAL is the transition from pixel-level reconstruction to semantic preservation. The framework consists of two primary components: Relaxed Cycle-Consistency and an Augmented Discriminator.
Relaxed Cycle-Consistency
Instead of the traditional constraint $x \approx G_{T \to S}(G_{S \to T}(x))$, ACAL introduces a "relaxed" version. It requires that a task-specific model (e.g., a pre-trained classifier) produces consistent outputs for both the original and the cycled images. For instance, if an image of a digit "7" is translated to the target domain and back, the final output must still be classified as a "7." This allows the generator to adapt the style freely while anchoring the semantic content necessary for the downstream task.
Augmented Discriminator
The task-specific model also assists the adversarial training process. By minimizing task loss on mapped samples, the framework encourages the generators to disentangle domain-specific "style" from task-relevant "content." This unified approach handles supervised, unsupervised, and semi-supervised domain adaptation within a single training algorithm.
Experimental Evaluation
The authors evaluated ACAL across two distinct modalities to demonstrate its versatility.
Visual Digit Adaptation
Experiments were conducted using various digit datasets, including MNIST, SVHN, USPS, and MNISTM. A particular focus was placed on "low-resource" scenarios, such as using only 10 labeled samples per class in the target domain.
Speech Recognition
The framework was also applied to the TIMIT dataset for acoustic domain adaptation. The task involved adapting a model trained primarily on male voices (source) to perform better on female voices (target), which were underrepresented in the training set.
Key Results and Discussion
ACAL demonstrated substantial improvements over baseline methods, especially in data-constrained environments.
- Low-Resource Performance: On the SVHN to MNIST task with 10 samples per class, ACAL achieved an accuracy of 93.9%, significantly outperforming standard CycleGAN (45.5%) and models trained solely on target data (79.2%).
- High-Resource Competitiveness: In several unsupervised settings, ACAL using only a few samples outperformed existing state-of-the-art models that utilized much larger datasets.
- Acoustic Adaptation: In the TIMIT speech task, ACAL reduced the Phoneme Error Rate (PER) for female speakers by 2% absolute, nearly reaching the performance of a model trained exclusively on the full female dataset.
The use of a pre-trained source task model acts as a powerful regularizer, preventing the target discriminator from overfitting to limited samples—a common failure mode in traditional adversarial domain adaptation. This stability is crucial when target data is scarce.
Conclusion
Augmented Cyclic Adversarial Learning (ACAL) addresses the limitations of rigid cycle-consistency in domain adaptation. By prioritizing semantic preservation over pixel-perfect reconstruction, ACAL provides a robust and stable framework for low-resource settings. Future work may explore extending this approach to purely unsupervised tasks using self-supervised task models or autoencoders.
Summary updated