Source-linked AI summary
Few-Shot Adversarial Domain Adaptation
Saeid Motiian, Quinn Jones, Seyed Mehdi Iranmanesh, Gianfranco Doretto
TL;DR
Few-shot supervised domain adaptation is difficult because target datasets are small, while standard unsupervised alignment requires many target samples and may not semantically align classes. FADA uses four-group source-target pairings with a multi-class domain-class discriminator to jointly encourage domain confusion and semantic alignment, achieving effective adaptation with very few labeled target samples and quickly converging as samples increase.
Problem
Few-shot supervised adaptation must align source and target distributions despite limited target data, while unsupervised methods require sufficiently large target datasets and may fail to align same-class samples across domains.
Method
FADA forms four pair groups encoding domain and class relationships, then adversarially trains a four-output domain-class discriminator alongside inference and prediction functions.
Results
FADA works with one labeled target sample per category, reaches comparable state-of-the-art accuracy with 10 labeled target samples, and shows clear promise as a state-of-the-art approach across experiments.
Takeaways & Limitations
Adversarial learning can be augmented for supervised few-shot adaptation, with semantic distribution alignment remaining effective and accuracy converging quickly as labeled target samples increase.
Takeaways & Limitations
The approach assumes covariate shift between source and target input distributions and is motivated by settings where the target dataset is small.
Abstract
from arXiv · showhide
This work provides a framework for addressing the problem of supervised domain adaptation with deep models. The main idea is to exploit adversarial learning to learn an embedded subspace that simultaneously maximizes the confusion between two domains while semantically aligning their embedding. The supervised setting becomes attractive especially when there are only a few target data samples that need to be labeled. In this few-shot learning scenario, alignment and separation of semantic probability distributions is difficult because of the lack of data. We found that by carefully designing a training scheme whereby the typical binary adversarial discriminator is augmented to distinguish between four different classes, it is possible to effectively address the supervised adaptation problem. In addition, the approach has a high speed of adaptation, i.e. it requires an extremely low number of labeled target training samples, even one per category can be effective. We then extensively compare this approach to the state of the art in domain adaptation in two experiments: one using datasets for handwritten digit recognition, and one using datasets for visual object recognition.
1 Introduction
Deep models perform well with abundant labeled data, but scarce target labels make straightforward training difficult. FADA adapts adversarial learning to supervised few-shot adaptation, using target labels and reporting effective gains even with very few samples.
- Motivation: Scarce labeled data limits straightforward deep-learning approaches for many computer-vision problems.Practitioners commonly reuse a model trained on a large source domain and adapt it to a smaller target domain.
- Motivation: Supervised domain adaptation can outperform unsupervised adaptation with the same amount of target data because it uses target labels.This makes SDA attractive when target data are scarce despite labeling costs.
- FADA: FADA applies adversarial learning to supervised adaptation when very few labeled target samples are available.The method is described as increasing performance rapidly as additional target samples are included, with even one sample potentially significant.
- FADA: The method creates four pair groups encoding source/target domains and same/different class labels, then trains a four-class discriminator adversarially.The discriminator is confused while source classification accuracy is maintained, encouraging domain confusion and semantic class alignment.
2 Related work
Prior domain-adaptation methods reduce dataset shift through domain mapping, shared latent spaces, or target regularization, while newer approaches use Siamese, coupled, and adversarial architectures. FADA extends this line with four-group pairing and a multi-class discriminator that incorporates class information.
- Domain adaptation: Dataset bias arises when image distributions differ between source and target datasets, making naive cross-dataset classification sub-optimal.
- Domain adaptation: Prior adaptation methods map domains, learn shared latent spaces, regularize target fitting, or combine these strategies.UDA can use domain mapping and shared representations, whereas SDA uses shared spaces, regularization, or both.
- Neural architectures: Siamese and coupled networks have been used to learn shared or related representations across source and target domains.Examples include soft label distribution matching, activation-correlation alignment, classifier-level fusion, and positive/negative pair construction.
- Adversarial adaptation: Adversarial learning has produced domain-adaptation approaches including coupled generative models and discriminative methods using unlabeled target data.
- FADA: FADA modifies pairing and discrimination by encoding domain and class labels in four pair groups rather than using a standard binary adversarial discriminator.Its multi-class discriminator supports a shared feature space while encouraging semantic alignment, which other adversarial UDA approaches do not address.
3 Few-shot adversarial domain adaptation
FADA addresses supervised domain adaptation with very limited labeled target data by combining domain confusion with semantic class alignment. It uses paired source–target samples and a four-class domain-class discriminator, with weight sharing to reduce overfitting.
- Problem: The SDA problem is to learn a target-effective predictor when source and target input distributions differ but labeled target data are scarce.The setting can include one labeled target sample per class or no target samples for some classes.
- Motivation: UDA can require many target samples and may align domains without ensuring same-class samples from different domains are nearby.This motivates supervised adaptation when target data are limited and class labels are available.
- Pair construction: FADA augments scarce target data by pairing each target sample with training source samples and forming four groups that encode pairwise class and domain information.The groups distinguish same-class source pairs, same-class cross-domain pairs, different-class source pairs, and different-class cross-domain pairs.
- Domain-class discrimination: A domain-class discriminator replaces the binary domain discriminator and predicts which of the four pair groups it receives.The discriminator uses the concatenated outputs of the appropriate inference functions and a softmax classifier trained with categorical cross-entropy.
- Adversarial objective: Adversarial updates make the discriminator confuse cross-domain pairs with same-domain pairs while preserving positive-versus-negative pair separability.The objective balances classification and confusion, thereby targeting both domain confusion and semantic class alignment.
- Model constraint: FADA uses weight sharing, gs = gt = g, because very few target samples are available.The paper describes weight sharing as a regularizer that can reduce overfitting, while noting that constrained or asymmetric inference functions are also possible.
4 Experiments
Experiments evaluate FADA on handwritten-digit and Office domain-adaptation tasks using very few labeled target samples. FADA performs effectively with scarce supervision, including one labeled target sample per category for digit adaptation, and achieves performance comparable to the state of the art on Office.
- The experiments cover MNIST, USPS, SVHN, and Office domain-adaptation datasets.The digit experiments consider six cross-domain tasks, while Office includes Amazon, Webcam, and DSLR domains.
- FADA works with one labeled target sample per category, and accuracy increases as the number of labeled target samples rises.The paper reports comparable state-of-the-art accuracies with 10 labeled target samples instead of more than thousands of unlabeled target samples.
- Fine-tuning improves the source-only lower bound but performs worse than supervised domain-adaptation methods in the reported digit experiments.The lower bound trains the base model using only source samples and no adaptation.
- Weight sharing improves the U → M result with n = 5 compared with training without shared weights.Without weight sharing, the average accuracy is 84.1 over 10 repetitions, below the weight-sharing case.
- FADA achieves performance comparable to the state of the art on the 31-class Office dataset.The Office comparison uses only three labeled target examples per category, while reported UDA methods use all target samples as unlabeled data, so the protocols are not exactly comparable.
5 Conclusions
The paper concludes that combining classification and adversarial losses effectively addresses supervised domain adaptation with very few labeled target samples. Accuracy converges quickly as more target samples become available, and the approach shows promise in the reported experiments.
- FADA combines classification and adversarial losses for supervised domain adaptation in the few-shot regime.
- Point-wise surrogates of distribution distances and similarities effectively address semantic distribution alignment with very few labeled target samples.
- SDA accuracy converges quickly as more labeled target samples per category become available.
- The approach is general because its architecture sub-components can be changed.
- The experiments show clear promise, including new state-of-the-art performance in the reported experiments.