MS-DIAL: Multi-Source Domain Alignment Layers for Unsupervised Domain Adaptation

Silva, Lucas Fernando Alvarenga e; Almeida, Jurandy

doi:10.5753/wvc.2020.13490

Cited by 4 publications

(4 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…On the other hand, OSDA aims to tackle two main challenges: handle irrelevant categories in the target domain that do not appear in the source domain (the OS problem) and deal with potential data distribution misalignment between the source and target domains (the UDA problem). Despite abundant research on OS [22]- [24] and UDA [4], [25]- [27], OSDA has been less explored.…”

Section: Related Workmentioning

confidence: 99%

“…Minor adjustments are then made through a jointly applied linear transformation on all distributions. In our study, we extend the applicability of DIAL to handle multiple source domains (M > 1), introducing an approach called MS-DIAL [25] that can be used in conjunction with any off-the-shelf MSDA methods.…”

Section: Uda: Domain Alignment Layersmentioning

confidence: 99%

“…In the initial stage, we discover that aligning domains in different feature spaces yields better results compared to solely performing domain alignment on the output representations of deep models. To enhance the transferability of DL models, we propose embedding MS-DIAL [25] to align the source and target distributions in various levels of feature spaces. These layers can be seamlessly integrated into the network backbones of existing MSDA methods, leading to a substantial improvement in performance, with relative gains of up to +30.64% observed in their classification accuracies.…”

Section: Osda: Dealing With Negativesmentioning

confidence: 99%

See 2 more Smart Citations

Open Set Domain Adaptation Methods in Deep Networks for Image Recognition

Alvarenga e Silva,

Almeida

2023

Anais Estendidos Da XXXVI Conference on Graphics, Patterns and Images (SIBRAPI Estendido 2023)

View full text Add to dashboard Cite

Deep learning (DL) has revolutionized various fields through its remarkable capacity to learn from raw data. However, in uncontrolled environments like in the wild, the performance of these systems might degrade to some extent, especially with unlabeled datasets. Naive approaches train DL models on labeled datasets (source domains) that resemble the unlabeled test dataset (target domain), but nonetheless, this approach may not yield optimal results due to domain and category-shift problems. These issues have been the primary focus of Unsupervised Domain Adaptation (UDA) and Open Set Recognition research areas. To address the domain-shift problem, we introduced the Multi-Source Domain Alignment Layers (MS-DIAL), a structural solution for multi-source UDA. MS-DIAL aligns the source domains and the target domain at various levels of the feature space, individually achieving competitive results comparable to the state-of-the-art, and when combined with other UDA methods, it further enhances transferability by up to 30.64% in relative performance gains. Subsequently, we tackled the demanding setup of Open Set Domain Adaptation (OSDA), where both domain and category-shift issues coexist. Our proposed approach involves dealing with negatives, extracting a high-confidence set of unknown instances, and using them as a hard constraint to refine the classification boundaries of OSDA methods. We assessed our proposal in an extensive set of experiments, which achieved up to 5.8% of absolute performance gains.

show abstract

Section: Related Workmentioning

confidence: 99%

Section: Uda: Domain Alignment Layersmentioning

confidence: 99%

Section: Osda: Dealing With Negativesmentioning

confidence: 99%

See 1 more Smart Citation

Open Set Domain Adaptation Methods in Deep Networks for Image Recognition

Alvarenga e Silva,

Almeida

2023

Anais Estendidos Da XXXVI Conference on Graphics, Patterns and Images (SIBRAPI Estendido 2023)

View full text Add to dashboard Cite

show abstract

“…OSDA mixes both OS and UDA, a more realistic scenario where irrelevant categories in the target domain do not appear in source domain and distributions of partially shared categories between source and target domains are not well aligned. Despite the vast literature on OS [5], [14], [18] and UDA [4], [15], [19], [20], OSDA has so far been little-studied.…”

Section: Related Workmentioning

confidence: 99%

Tightening Classification Boundaries in Open Set Domain Adaptation through Unknown Exploitation

e Silva,

Sebe,

Almeida

2023

2023 36th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI)

Self Cite

View full text Add to dashboard Cite

Convolutional Neural Networks (CNNs) have brought revolutionary advances to many research areas due to their capacity of learning from raw data. However, when those methods are applied to non-controllable environments, many different factors can degrade the model's expected performance, such as unlabeled datasets with different levels of domain shift and category shift. Particularly, when both issues occur at the same time, we tackle this challenging setup as Open Set Domain Adaptation (OSDA) problem. In general, existing OSDA approaches focus their efforts only on aligning known classes or, if they already extract possible negative instances, use them as a new category learned with supervision during the course of training. We propose a novel way to improve OSDA approaches by extracting a high-confidence set of unknown instances and using it as a hard constraint to tighten the classification boundaries of OSDA methods. Especially, we adopt a new loss constraint evaluated in three different means, (1) directly with the pristine negative instances; (2) with randomly transformed negatives using data augmentation techniques; and(3) with synthetically generated negatives containing adversarial features. We assessed all approaches in an extensive set of experiments based on OVANet, where we could observe consistent improvements for two public benchmarks, the Office-31 and Office-Home datasets, yielding absolute gains of up to 1.3% for both Accuracy and H-Score on Office-31 and 5.8% for Accuracy and 4.7% for H-Score on Office-Home.

show abstract