PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Inheritable Models for Open-Set Domain Adaptation

Jogendra Nath Kundu, Naveen Venkat, Ambareesh Revanur, Rahul M V, R. Venkatesh Babu

arXiv:2004.04388v1cs.CVcs.LG

TL;DR

Open-set domain adaptation must recognize target categories absent from labeled source data, but existing approaches rely on source and target data being available together. This paper proposes inheritable source models and an objective inheritability measure for vendor-client adaptation without data exchange; it reports state-of-the-art open-set DA performance and shows that higher measured inheritability tracks better adaptation on a target domain.

  • Problem

    Existing open-set DA requires labeled source data alongside unlabeled target instances, a constraint when privacy, proprietary, or copyright concerns prevent data sharing.

    Source: §1 · pp. 1–2

  • Method

    The paper formalizes inheritable source models and an objective inheritability measure so clients can adapt using the model and unlabeled target data without accessing source data.

    Source: §1 · p. 2

  • Results

    The authors report state-of-the-art open-set DA performance, and on Office-31 adaptation to W was better for source D, whose model had higher inheritability, than for source A.

    Source: §5.3 · p. 7, §6 · p. 8

  • Takeaways & Limitations

    Measuring inheritability can help a client select a suitable vendor model for a target domain without requiring access to the vendor’s source data.

    Source: §1 · p. 2, §5.3 · p. 7

    The confidence-based weighting relies on the hypothesis that source instances receive the highest source-class confidence, followed by target-shared instances and then target-unknown instances.

    Source: §4.2 · p. 5

Abstract

from arXiv · show

There has been a tremendous progress in Domain Adaptation (DA) for visual recognition tasks. Particularly, open-set DA has gained considerable attention wherein the target domain contains additional unseen categories. Existing open-set DA approaches demand access to a labeled source dataset along with unlabeled target instances. However, this reliance on co-existing source and target data is highly impractical in scenarios where data-sharing is restricted due to its proprietary nature or privacy concerns. Addressing this, we introduce a practical DA paradigm where a source-trained model is used to facilitate adaptation in the absence of the source dataset in future. To this end, we formalize knowledge inheritability as a novel concept and propose a simple yet effective solution to realize inheritable models suitable for the above practical paradigm. Further, we present an objective way to quantify inheritability to enable the selection of the most suitable source model for a given target domain, even in the absence of the source data. We provide theoretical insights followed by a thorough empirical evaluation demonstrating state-of-the-art open-set domain adaptation performance.

1. Introduction

Open-set DA must handle target-only categories, but existing methods rely on coexisting source and target data. The paper introduces a vendor-client paradigm with inheritable models and an inheritability measure for source-free model choice.

1. Introduction

  • Domain shift can cause test-time performance drops when source data does not characterize the target environment.The introduction situates this challenge in visual-recognition tasks with limited available source datasets.Source: §1 · p. 1
  • Open-set DA must handle target categories absent from the source, while avoiding negative transfer from those unshared categories.Direct distribution alignment can reduce performance through interference from unshared categories.Source: §1 · pp. 1–2
  • Existing UDA methods assume labeled source data during adaptation, limiting their suitability where proprietary, sensitive, or private data cannot be shared.The paper identifies privacy and copyright constraints affecting both vendors and clients.Source: §1 · p. 2
  • Inheritable models transfer task-specific knowledge from a source-trained model so clients can adapt without accessing source-domain data.The inheritability measure helps clients choose a suitable model for a specific target domain.Source: §1 · p. 2
  • The vendor-client paradigm enables open-set adaptation without coexisting source and target domains or data exchange.The paper proposes inheritable models and an objective inheritability measure as part of this paradigm.Source: §1 · p. 2
  • The authors report state-of-the-art open-set DA performance using inheritable models.The paper also presents theoretical insights and empirical evaluation.Source: §1 · p. 2

2. Related Work

Prior DA approaches address shared-label alignment, open-set recognition, domain generalization, or data-free distillation, but the paper distinguishes its setting by combining domain and category shift without data exchange.

Closed-set DA

  • Closed-set DA methods align domains under a shared-label assumption, leaving them prone to negative transfer when target-only categories occur.The cited approaches include discrepancy measures, adversarial feature matching, and domain-specific normalization.Source: §2 · p. 2

Open-set DA

  • Open-set DA methods identify or reject unknown target instances, but the described approaches still assume coexisting source and target data.The passage notes hyperparameter sensitivity and negative-transfer concerns for some methods, while STA uses instance-level weights to mitigate negative transfer.Source: §2 · p. 2

Domain Generalization

  • Domain-generalization methods rely on multiple coexisting source domains with shared label sets, unlike this paper’s single-source open-set setting.The passage identifies category shift among vendors’ available data as a practical issue for those methods.Source: §2 · p. 2

Data-free Knowledge Distillation

  • Data-free knowledge distillation transfers teacher outputs without training data, but the described methods do not address domain shift or unknown target categories.The passage says teacher domain bias can pass to the student and that these methods are not designed for recognizing new target categories.Source: §2 · p. 2

Related-work distinction

  • The paper presents its formulation as robust to negative transfer while enabling domain adaptation without data exchange.This distinguishes it from the prior approaches discussed in the section.Source: §2 · p. 2

3. Unsupervised Open-Set Domain Adaptation

The section formalizes unsupervised open-set domain adaptation and the vendor-client setting, then defines when a source model is inheritable for learning a target predictor without source data.

3. Unsupervised Open-Set Domain Adaptation

  • The section introduces formal definitions for the vendor-client paradigm and inheritability in unsupervised open-set domain adaptation.These definitions establish the setting and conditions used in the following design discussion.Source: §3 · p. 3, §3.2 · p. 3

3.1. Preliminaries

  • UODA assigns source-shared target instances to shared classes and target-only instances to an unknown label, with performance measured by target-predictor error.The setup uses labeled source data and unlabeled target instances.Source: §3.1 · p. 3

3.2. The vendor-client paradigm

  • In the vendor-client paradigm, a vendor trains a source predictor on labeled source data and shares the model for client adaptation using unlabeled target instances.The paradigm assumes no data exchange and allows one vendor model to be shared with multiple clients, reducing repeated source-training effort.Source: §3.2 · p. 3

3.3. Inheritability

  • A source predictor is inheritable if unlabeled-target adaptation can learn a target predictor whose error is within ε of the best predictor in H with probability at least 1−δ.This formalizes reliable transfer of task-specific knowledge without access to source data.Source: §3.3 · p. 3

4. Approach

The approach equips a source model to retain class-separability knowledge and characterize OOD regions, then uses confidence-based inheritance and selective alignment for target adaptation. Its theoretical discussion relates pseudo-labeling to the inheritability criterion for shared classes.

4. Approach

  • Open-set adaptation must preserve class-separability while detecting new target categories as unknown to avoid negative-transfer.The paper demonstrates this design for UODA.Source: §4 · p. 3, §4.1 · p. 3

4.1. Vendor trains an inheritable model

  • The vendor model combines a source classifier for class-separability with an auxiliary OOD classifier for the negative region.Joint training lets the model capture source-class structure while characterizing the negative region.Source: §4.1 · pp. 4–5
  • Feature-splicing replaces top-percentile activations with those from another class, creating negative features that are clustered into labels for OOD training.The paper uses these synthetic instances to train the auxiliary classifier to distinguish negative classes.Source: §4.1 · p. 4
  • Vendor training first pretrains the feature extractor and source classifier, then freezes the backbone and trains the remaining modules using source and synthetic-negative examples.The second stage includes the source and auxiliary OOD classifiers.Source: §4.1 · p. 4

4.2. Client adapts to the target domain

  • The model’s inheritability measure uses source-classifier confidence: source instances are expected to score highest, shared target instances next, and unknown instances lowest.The dataset-level inheritability aggregates this instance-level measure; higher values imply smaller domain shift and greater inheritability.Source: §4.2 · p. 5
  • Adaptation pseudo-labels the highest-inheritability target instances, while soft instance weights guide likely shared samples toward source-density regions and lower-weight samples toward negative regions.The combined tuning loss selectively aligns shared classes while avoiding negative-transfer.Source: §4.2 · pp. 5–6

4.3. Theoretical Insights

  • Pseudo-labeling is presented as conditioning the target model to satisfy the shared-class inheritability criterion; improving prediction precision during adaptation is described as tightening its theoretical bound.The paper relates increasing correct labels to a larger effective sample size in the bound.Source: §4.3 · p. 6

5. Experiments

Experiments evaluate open-set domain adaptation on Office-31, Office-Home, and VisDA, using ResNet-50 for the first two and VGG-16 for VisDA. Results compare inheritable-model adaptation with prior methods and examine sensitivity, inheritability, and efficiency.

5.2. Results

  • Inheritable-model adaptation outperforms prior methods across the reported benchmarks, although the supplied passages provide no table values.The comparison covers Office-31, Office-Home, and VisDA; the other methods use source-domain data during adaptation.Source: §5.2 · p. 6, Table 1 · p. 7, Table 2 · p. 7
  • Performance shows low sensitivity to vendor-training K and d, improves as k increases, and remains strong across a range of openness settings.The k=0 ablation does not enforce Linh; the authors interpret the increase with k as supporting inheritance of class-separability knowledge.Source: §5.2 · p. 7, Figure 4 · p. 7, Figure 5 · p. 8
  • For target-shared instances, Proxy A-distance is lower than for a source-trained ResNet-50, while it is higher for target-unknown instances.The authors interpret this pattern as shared-distribution alignment alongside separation of target-unknown instances.Source: §5.2 · p. 7, Figure 6 · p. 8

5.3. Discussion

  • For target W, the model trained on D has higher inheritability than the model trained on A, and D→W performs better than A→W.The authors propose measuring inheritability to help a client choose a vendor model without accessing its source data.Source: §5.3 · p. 7, Figure 6 · p. 8
  • On A→D, the instance-level inheritability histogram compares target-shared with target-unknown samples; top-15-percentile confidence predictions are near 1 precision at epoch 0.As adaptation proceeds, the top-confidence precision rises, tightening the bound in Eq. 12.Source: §5.3 · p. 8, Figure 5 · p. 8
  • On A→D, t-SNE visualizations show the method performing equally well despite adaptation without source data.The comparison is with STA, which uses source data during adaptation.Source: §5.3 · p. 8, Figure 6 · p. 8
  • For one vendor and two clients, the reported total cost is 435s for the inheritable-model pipeline versus 1150s for STA.The inheritable-model total includes 250s of vendor training and 138s of client adaptation; STA adaptation averages 575s per client.Source: §5.3 · p. 8

6. Conclusion

The paper proposes inheritable models for open-set domain adaptation without co-existing source and target domains, and an objective inheritability measure for selecting a source model without source-data access. Empirical evaluation reports state-of-the-art performance; extending the approach to multiple vendors and clients is future work.

6. Conclusion

  • The paper introduces a vendor-client paradigm for open-set domain adaptation without co-existing source and target domains.It proposes inheritable models to support this setting.Source: §6 · p. 8
  • An objective inheritability measure is proposed to select a suitable source model for a target domain without accessing source data.Source: §6 · p. 8
  • The authors report state-of-the-art open-set domain adaptation performance and identify multiple vendors and clients as future work.Source: §6 · p. 8

Supplementary: Towards Inheritable Models for Open-Set Domain Adaptation

The supplementary material lists a theoretical derivation, implementation details, alternate out-of-distribution sample-generation methods, and additional experimental details.

Supplementary: Towards Inheritable Models for Open-Set Domain Adaptation

1. Derivation of Result 1

The derivation specializes a general domain-adaptation bound to target-only learning over target-shared instances. It uses this bound to formalize how class-separability can be inherited without source data.

1. Derivation of Result 1

  • Setting α = 1 and β = 1 specializes the general bound to target-only adaptation, since the objective is target performance and source data are unavailable.The derivation takes a limit as α and β approach 1 to handle an indeterminate form.Source: §1 · p. 11
  • The client selects top-k target instances by inheritability value w; their pseudo-labels are treated as target-shared samples with small label noise.Minimizing Linh searches the hypothesis space for an empirical minimizer on these instances.Source: §1 · p. 11
  • With probability at least 1−δ, Result 1 gives a target-shared sample guarantee whose error term ε depends on VC dimension d, sample size m, and δ.The stated setup uses a labeled sample drawn from the target-shared distribution.Source: §1 · p. 12
  • Minimizing Linh prunes the hypothesis space to satisfy the inheritability criterion, making class-separability knowledge inheritable for adaptation.As Ltune improves target-model precision, the number of correctly labeled target-shared instances grows and tightens the bound.Source: §1 · p. 12

2. Algorithm and implementation details

The section outlines vendor-side model preparation and client-side adaptation, including the architecture and feature-splicing procedure. The client adapts using unlabeled target data and the vendor’s inheritable model.

2.1. Architecture

  • Office-31 and Office-Home use ResNet-50 features of dimension 2048, while VisDA uses pooled VGG-16 features of dimension 512.The E and G modules use fully connected and batch-normalization layers with ELU nonlinearity.Source: §2.1 · p. 12

2. Algorithm and implementation details

  • The workflow pretrains a source model, trains it as inheritable, then adapts a client model using target data and the vendor model.The pseudocode separates vendor and client stages.Source: §2.2 · p. 12, §2 · p. 12, §2.3 · p. 13

2.2. Vendor trains an inheritable model

  • The vendor pretrains on labeled source data, then splices top-dimension activations from different-class features to create negative instances for OOD training.K-means assigns the generated instances negative-class labels before the vendor trains the model with source and negative instances.Source: §2.2 · pp. 12–13

2.3. Client adapts to the target domain

  • The client selects high-inheritability target instances for pseudo-labeling, then adapts the model on those instances and unlabeled target batches by minimizing Linh + Ltune.The algorithm normalizes unlabeled-instance weights by the maximum weight in each batch.Source: §2.3 · p. 13

3. Alternate methods to generate OOD samples

The section compares two alternative ways to generate OOD features with feature-splicing. Linear interpolation performs worse, while random suppression performs similarly but requires tuning a suppression range.

3. Alternate methods to generate OOD samples

  • Linear interpolation performs worse than feature-splicing, which can mimic plausible OOD samples by suppressing class-specific traits.The interpolated features occupy a constrained region between source classes.Source: Table 2 · p. 14, §3 · p. 14
  • Random suppression performs similarly to feature-splicing, but requires searching for a suppression scale range of 0.2 to 0.3.Feature-splicing instead chooses replacements for class-discriminative activations for each instance.Source: Table 2 · p. 14, §3 · p. 14

4. Miscellaneous

The section reports training-time comparisons and describes the Office-31, Office-Home, and VisDA benchmarks. Vendor-side source training is reused across clients in the proposed approach.

4.1. Training time comparison

  • 250s of vendor-side source training supports multiple clients, each adapting in 69s on average; STA requires source data during each client’s adaptation.On identical settings, STA averaged 177s for Step 1 and 398s for Step 2 on A→D and A→W.Source: §4.1 · p. 14

4.2. Dataset description

  • The experiments use Office-31 (4,652 images; three domains), Office-Home (about 15,500 images; 65 classes and four domains), and VisDA (about 150k synthetic and 56k real images).Office-Home has higher domain shift than Office-31; each benchmark specifies shared and target-unknown classes.Source: §4.2 · pp. 14–15
Loading 2004.04388v1…