PaperlayerView exact revision on arXiv

Source-linked AI summary

A Survey on Negative Transfer

Wen Zhang, Lingfei Deng, Lei Zhang, Dongrui Wu

arXiv:2009.00909v4cs.LGcs.CVstat.ML

TL;DR

Negative transfer can reduce target-domain performance when transfer-learning assumptions are unmet, yet no comprehensive survey had organized its formulation, causes, and remedies. This paper surveys about fifty representative approaches across four categories and discusses related fields, providing the first comprehensive survey on negative transfer.

  • Problem

    Negative transfer remains a challenging transfer-learning problem, and no comprehensive survey had systematically synthesized its definition, contributing factors, and mitigation approaches.

  • Method

    The paper defines negative transfer and its factors, then systematically reviews about fifty representative approaches across secure transfer, similarity estimation, distant transfer, and mitigation.

  • Results

    The survey categorizes and reviews about fifty negative-transfer approaches and introduces related concepts and fields, constituting the first comprehensive survey on negative transfer.

  • Takeaways & Limitations

    Negative-transfer research can be examined through four perspectives: secure transfer, domain similarity estimation, distant transfer, and negative-transfer mitigation.

  • Takeaways & Limitations

    A-distance may perform poorly because it ignores label-space differences and considers only marginal distribution discrepancies.

Abstract

from arXiv · show

Transfer learning (TL) utilizes data or knowledge from one or more source domains to facilitate the learning in a target domain. It is particularly useful when the target domain has very few or no labeled data, due to annotation expense, privacy concerns, etc. Unfortunately, the effectiveness of TL is not always guaranteed. Negative transfer (NT), i.e., leveraging source domain data/knowledge undesirably reduces the learning performance in the target domain, has been a long-standing and challenging problem in TL. Various approaches have been proposed in the literature to handle it. However, there does not exist a systematic survey on the formulation of NT, the factors leading to NT, and the algorithms that mitigate NT. This paper fills this gap, by first introducing the definition of NT and its factors, then reviewing about fifty representative approaches for overcoming NT, according to four categories: secure transfer, domain similarity estimation, distant transfer, and NT mitigation. NT in related fields, e.g., multi-task learning, lifelong learning, and adversarial attacks, are also discussed.

1 INTRODUCTION

Transfer learning addresses distribution shifts and limited target-domain data by leveraging related source-domain data or knowledge, but violations of its assumptions can cause negative transfer. This survey defines NT, analyzes its causes, and systematically reviews about fifty approaches for reliable transfer, focusing on homogeneous closed-set classification.

  • Motivation: Real-world distribution shifts can make models generalize poorly, while collecting matching labeled data is often costly or constrained by privacy concerns.Examples include different camera resolutions or lighting conditions and individual differences in brain-computer interfaces.
  • Transfer Learning: Transfer learning leverages data or knowledge from related source domains to improve learning and generalization in a new target domain.The target domain often has limited labeled data, motivating transfer from available related domains.
  • Negative Transfer: Violating assumptions about task relatedness, distribution similarity, or model suitability can cause negative transfer, where source information decreases target-domain performance.NT is identified as a long-standing and challenging problem in transfer learning.
  • Reliable Transfer: Reliable transfer requires deciding what, how, and when to transfer, although most prior research focused only on what and how.The survey notes that considering all three questions is necessary to avoid NT.
  • Survey Contributions: The paper fills a survey gap by reviewing about fifty representative NT approaches, defining NT and its factors, and covering secure transfer, similarity estimation, distant transfer, and mitigation.Its main scope is homogeneous, closed-set classification, with unchanged source and target tasks and unchanged target feature and label spaces during testing.

2 BACKGROUND KNOWLEDGE

This section defines transfer learning and negative transfer, introduces four transfer-learning categories, and identifies domain divergence and transfer-algorithm design among factors contributing to negative transfer. It also notes that the negative-transfer gap may be uncomputable without labeled target data.

  • Transfer learning: Transfer learning uses source and target-domain information to learn a target-domain hypothesis with small expected target loss.The target loss is defined as ϵT(h) = Ex,y∼PT(X,Y)[ℓ(h(x), y)].
  • Transfer learning: Transfer learning approaches are categorized as instance based, feature based, model/parameter based, and relation based.These categories respectively involve weighted source samples, shared representations, parameters, and preserved internal relationships or rules.
  • Negative transfer: Negative transfer occurs when using source information increases target-domain test error relative to the same algorithm without source information.Its degree is evaluated by the negative transfer gap, or NTG.
  • Negative transfer: The negative transfer gap is not always computable because unsupervised transfer lacks labeled target data needed to measure the no-source target error.In that scenario, ϵT(θ(∅, T)) is impossible to compute.
  • Factors of negative transfer: Four factors can contribute to negative transfer, with domain divergence described as its root and transfer algorithms requiring security guarantees or careful design.Methods that do not explicitly minimize source–target divergence are more likely to produce negative transfer.

3 RELIABLE TL

Reliable transfer learning organizes strategies for alleviating or avoiding negative transfer around secure transfer, domain similarity estimation, distant transfer, and negative-transfer mitigation. Estimated similarity guides whether to refuse transfer, mitigate negative transfer, or directly combine data, while also enabling transferable-source selection.

  • Reliable TL: Secure transfer can overcome negative transfer with theoretical guarantees, regardless of source–target domain similarity.Most other approaches assume some source–target similarity.
  • Reliable TL: Accurately estimating source–target domain similarity helps determine which strategy should handle negative transfer.The proposed reliable TL scheme considers existing strategies for alleviating or avoiding negative transfer.
  • Reliable TL: Low similarity supports refusing transfer or using distant transfer; medium similarity supports negative-transfer mitigation; high similarity supports direct data concatenation and classifier training.These choices are made using the estimated domain similarity.
  • Reliable TL: With multiple source domains, domain similarity estimation can select the most transferable sources.This extends similarity-based strategy selection beyond a single source domain.

4 SECURE TRANSFER

Secure transfer designs objectives that explicitly avoid negative transfer, ensuring transferred models outperform or do no worse than non-transfer baselines. Approaches include similarity-adaptive Bayesian transfer, worst-case guarantees, safe regularization, and error-bound estimation.

  • Secure transfer: Secure transfer explicitly requires transfer to outperform learning without transfer, making it useful when domain similarity is difficult to estimate.The section notes that only a few secure-transfer approaches exist, with some limited to classification and others to regression.
  • Bayesian adaptive transfer: Bayesian adaptive learning adjusts the transfer schema using estimated source–target dissimilarity, sharing less information when domain correlation is low.When λ is close to 0, only parameters in the transfer kernel may be shared.
  • Deep adaptation: Deep face-detector adaptation avoids negative transfer and catastrophic forgetting by constraining the target detector so its relative performance loss is non-positive.Thus, the optimized target detector is always no worse than the pretrained source detector.
  • Safe weakly supervised learning: SAFEW avoids negative transfer by optimizing the worst-case performance gain over a baseline trained only on labeled target data.Its target hypothesis combines multiple source-domain base learners under nonnegative weights summing to one.

5 DOMAIN SIMILARITY ESTIMATION

Domain similarity estimation is central to reliable transfer learning and comprises feature-statistics-, test-performance-, and fine-tuning-based approaches. These methods differ in their data requirements and ability to capture marginal versus joint distribution differences, while newer measures support source selection without source data.

  • Overview: Domain similarity estimation methods fall into three groups: feature statistics based, test performance based, and fine-tuning based.The three groups are summarized as the main taxonomy for domain similarity estimation.
  • Feature statistics based: Feature-statistics methods measure distribution discrepancy using representations or statistics such as means and covariances, with MMD offering a simple, effective nonparametric measure.Other measures include correlation coefficient, KL-divergence, HSIC, Bregman divergence, optimal transport, and Wasserstein distance.
  • Feature statistics based: MMD and A-distance primarily assess marginal discrepancy, so they can fail when conditional distributions or label spaces differ substantially.A-distance is NP-hard to compute exactly and is commonly approximated with a linear domain classifier.
  • Test performance based: Test-performance methods select or rank sources using labeled target performance, classifier correlation, or domain discrimination, but may require labeled target samples or an accessible target model.STIL removes historical models whose Q-statistics are close to 0, thereby addressing negative transfer.
  • Fine-tuning based: Label-informed divergence improves A-distance by measuring joint-distribution shift in time-evolving target domains.Fine-tuning-based measures include mean silhouette coefficient, NCE, LEEP, and TransRate; TransRate was reported to select source data, architectures, and network layers effectively.

6 DISTANT TRANSFER

Distant transfer addresses negative transfer caused by very low source–target similarity by bridging domains through one or more intermediate domains. Reviewed methods select useful source instances, learn shared intermediate representations, and incorporate task-related side information, including for scarce-data applications such as coronavirus diagnosis.

  • 6 DISTANT TRANSFER: Distant transfer bridges dramatically different source and target domains through one or more intermediate domains to address negative transfer from low data similarity.The passage contrasts text-to-image transfer as a likely cause of negative transfer and presents distant or transitive transfer as a solution.
  • 6 DISTANT TRANSFER: Tan et al. selected useful source instances, constructed multiple intermediate domains, and jointly minimized reconstruction errors on selected intermediate-domain instances and all target instances.They learned paired encoding and decoding functions for this reconstruction objective.
  • 6 DISTANT TRANSFER: The method used an auto-encoder reconstruction framework with source and intermediate selection indicators, regularization, and side information such as intermediate-domain predictions.Side information helped the model learn more task-related feature representations.
  • 6 DISTANT TRANSFER: Distant transfer has also been applied to scarce-data settings including medical diagnostics and remote sensing.These applications extend the strategy beyond the general formulation described for intermediate-domain transfer.
  • 6 DISTANT TRANSFER: Niu et al. transferred knowledge from object-recognition datasets and chest X-ray images to coronavirus diagnosis using convolutional auto-encoders that reconstruct domains in a shared intermediate feature space.The provided passage states that the method reconstructs common and medical image domains in the same intermediate feature space.

7 NT MITIGATION

NT mitigation enhances transferability at the domain, instance, feature, and model levels. Representative strategies select or weight transferable domains and instances, learn shared feature spaces while preserving discriminability, suppress noise, and adapt model normalization.

  • NT mitigation: Overall, NT mitigation combines domain, instance, feature, and model transferability enhancement to exploit source-target similarity without indiscriminate transfer.The surveyed framework organizes these approaches around data transferability enhancement, model transferability enhancement, and target prediction enhancement.
  • Data transferability enhancement: Domain selection or weighting aggregates source domains according to similarity, often outperforming simple averaging and, with learned weights, performing at least as well as the single best source model.Methods use MMD-based proximity, attention-based domain recognition, or optimized weights from source-model predictions on unlabeled target data.
  • Data transferability enhancement: Instance selection and weighting remove irrelevant or less transferable source data using pseudo-label confidence, clustering, active learning, or discriminator-estimated distribution density ratios.Active transfer learning selects class-balanced source samples similar to target samples, while discriminator gates assign source-sample weights w(x_i, y_i) = D(x_i, y_i)/(1 − D(x_i, y_i)).
  • Data transferability enhancement: Feature-level methods learn common latent spaces or enhance feature transferability, while jointly preserving discriminability to avoid poor classification.Approaches include dual transfer, singular-value regularization, batch spectral penalization, and sparse low-rank alignment with noise suppression.
  • Model transferability enhancement: Model transferability enhancement includes transferable normalization, which reduces batch-normalization domain shift and adapts each channel using distance-based transferability probabilities.TransNorm is applied after convolutional layers and uses source-target means and variances to quantify domain distance.

8 NT IN RELATED FIELDS

Negative transfer has been studied in multi-task learning, lifelong learning, and adversarial attacks. These fields expose task interference, forgetting, harmful forward transfer, and attacks on models or data as related challenges.

  • Multi-task learning: In multi-task learning, conflicting gradients among tasks may induce negative transfer, also called negative interference.Remedies include directly altering gradients and other techniques.
  • Multi-task learning: Multilingual models show negative transfer, especially for high-resource languages, with proposed remedies including parameter soft-sharing, meta-learning, and gradient vaccine.These models can process tens or even hundreds of languages simultaneously.
  • Lifelong learning: Lifelong learning faces catastrophic forgetting and harmful transfer from early tasks to later tasks, while forward negative transfer remains less investigated.Existing mitigation methods mainly address catastrophic forgetting through regularization, memory replay, and parameter isolation.
  • Adversarial attacks: In adversarial attacks, transferable adversarial examples and unsecured source data or models can degrade target learning performance and cause negative transfer.Evasion attacks can affect white-box teacher models, black-box student models, and transfer-learning parameters.

9 METHOD COMPARISON

The paper compares four categories of approaches for overcoming negative transfer: secure transfer, domain similarity estimation, distant transfer, and negative-transfer mitigation. Most existing work emphasizes negative-transfer mitigation and domain similarity estimation, with mitigation focused mainly on enhancing data transferability.

  • Method comparison: Most current negative-transfer research focuses on negative-transfer mitigation and domain similarity estimation.The comparison identifies these as the dominant categories in the literature.
  • Method comparison: Negative-transfer mitigation research mainly targets data transferability enhancement.This focus characterizes the primary objective within the mitigation category.
  • Method comparison: Approaches to overcoming negative transfer span many transfer-learning categories, indicating extensive attention across transfer learning.The compared categories are secure transfer, domain similarity estimation, distant transfer, and negative-transfer mitigation.
  • Method comparison: Most secure-transfer strategies are based on model adaptation.Model adaptation is the predominant basis reported for secure-transfer methods.
  • Method comparison: Most negative-transfer mitigation approaches consider two or more negative-transfer factors.The comparison highlights multi-factor consideration as a common design characteristic of mitigation methods.

10 CONCLUSIONS AND FUTURE RESEARCH

The paper systematically reviews about fifty approaches for handling negative transfer across four perspectives. It recommends strategy selection based on domain-similarity estimates and identifies future work in secure transfer, regression mitigation, and challenging open environments.

  • Conclusions: The paper categorizes about fifty representative negative-transfer approaches into secure transfer, domain similarity estimation, distant transfer, and negative-transfer mitigation.These categories provide the paper’s systematic framework for reviewing methods that address negative transfer in transfer learning.
  • Guidelines for coping with NT: When domain similarity is high, directly concatenate source and target data; when medium, use negative-transfer mitigation; when low, use distant transfer or no transfer.The recommended strategy changes according to the estimated similarity between source and target domains.
  • Guidelines for coping with NT: When domain similarity cannot be estimated, secure transfer may be used.This guideline applies when similarity-based strategy selection is unavailable.
  • Future research: Future work should develop secure transfer for unsupervised DA, few shot TL, and adversarial TL, and extend negative-transfer mitigation to regression problems.Most existing negative-transfer mitigation approaches currently target classification problems.
  • Future research: Future research should ensure positive transfer in open environments with continual data streams, heterogeneous features, private sources, unclear domain boundaries, and unseen or unknown categories.These conditions represent challenging open environments identified for future investigation.
Loading 2009.00909v4…