Source-linked AI summary
Characterizing and Avoiding Negative Transfer
Zirui Wang, Zihang Dai, Barnabás Póczos, Jaime Carbonell
TL;DR
Negative transfer can hurt target performance when source and target domains are weakly related, yet it has lacked a rigorous definition and systematic treatment. The paper formalizes the phenomenon, analyzes its factors, and introduces an adversarial discriminator gate to filter harmful source data. Across six deep transfer methods and four benchmark datasets, the approach consistently improves baseline performance and largely avoids negative transfer, including with degenerate source data.
Problem
Negative transfer is an insufficiently defined and analyzed failure mode in which weakly related source data can hinder target performance.
Method
The paper defines negative transfer formally, analyzes three factors, and uses an adversarial discriminator gate to filter potentially harmful source data.
Results
The method consistently improves six deep transfer baselines across four benchmark datasets and largely avoids negative transfer, even with degenerate source data.
Takeaways & Limitations
Negative transfer can be analyzed through algorithm, domain-divergence, and target-data factors, while discriminator gating provides a generic avoidance technique.
Takeaways & Limitations
The domain-adversarial approach assumes a feature space shared by source and target domains that remains discriminative for predicting outputs.
Abstract
from arXiv · showhide
When labeled data is scarce for a specific target task, transfer learning often offers an effective solution by utilizing data from a related source task. However, when transferring knowledge from a less related source, it may inversely hurt the target performance, a phenomenon known as negative transfer. Despite its pervasiveness, negative transfer is usually described in an informal manner, lacking rigorous definition, careful analysis, or systematic treatment. This paper proposes a formal definition of negative transfer and analyzes three important aspects thereof. Stemming from this analysis, a novel technique is proposed to circumvent negative transfer by filtering out unrelated source data. Based on adversarial networks, the technique is highly generic and can be applied to a wide range of transfer learning algorithms. The proposed approach is evaluated on six state-of-the-art deep transfer methods via experiments on four benchmark datasets with varying levels of difficulty. Empirically, the proposed method consistently improves the performance of all baseline methods and largely avoids negative transfer, even when the source data is degenerate.
1. Introduction
Transfer learning addresses labeled-data scarcity by reusing source-domain data, but weak source–target relatedness can cause negative transfer. The paper formalizes this problem, analyzes its factors, and proposes a generic adversarial approach to filter harmful source data.
- Deep neural networks often need substantial labeled data and assume training and test data share an underlying distribution.
- Negative transfer occurs when transferring from a weakly related source hinders target performance.
- Prior work left negative transfer insufficiently analyzed, including its definition, causes, prediction, and detection or avoidance with limited target labels.
- The paper defines negative transfer formally, identifies three underlying factors, and develops an adversarial-network technique that filters potentially harmful source data.
- Experiments across eight transfer methods and four benchmark datasets reveal the three factors and show improvements across six deep transfer methods under potential negative-transfer conditions.
2. Related Work
Transfer-learning research progressed from conventional instance weighting and feature mapping to deep methods that learn domain-invariant representations or reduce distribution mismatch. Negative transfer has been observed across several settings, but the literature lacks in-depth analysis.
- Early transfer methods used instance weighting and feature mapping, whereas deep transfer networks seek domain-invariant factors.
- Deep transfer methods have used MMD-based distribution matching and GAN-based subnetworks for transfer.
- Negative effects have been observed in simple classifiers, multi-source transfer, imbalanced distributions, and partial transfer learning.
- Despite growing attention to detecting and avoiding negative transfer, prior literature lacks in-depth analysis.
3. Rethink Negative Transfer
The paper reframes negative transfer as an algorithm-specific increase over a target-only baseline and identifies distribution divergence and labeled-target-data size as key factors. These insights motivate systematic avoidance of harmful source information.
- Transfer learning takes source and target data as input to produce a target hypothesis intended to outperform training on target data alone.
- The analysis focuses on a specific algorithm because increased risk may reflect algorithm differences rather than using source data.
- The negative transfer gap is R_PT(A(S,T)) − R_PT(A(∅,T)); negative transfer occurs when this gap is positive.
- Distribution shift between source and target joint distributions is identified as the root of negative transfer when algorithms fail to discard divergent information.
- The amount of labeled target data has a mixed effect because it changes both the target-only baseline and the ability to discover shared joint-distribution regularities.
- Without labeled target data, transfer must rely on marginal similarity, whereas considerable labeled target data can make shared-regularity discovery manageable.
4. Proposed Method
The method addresses DANN’s assumption that every source sample is useful by estimating source–target relevance and gating harmful samples, while augmenting adversarial learning with joint-distribution matching. It is designed to generalize beyond DANN to multiple adversarial and non-adversarial transfer methods.
- Motivation: DANN assumes every source sample can provide meaningful transfer knowledge, which can make unrelated samples harm target performance.This motivates selectively reducing the influence of harmful source samples.
- Architecture: The method’s architecture uses extracted features, predicted class and domain labels, classification loss, adversarial loss, a gradient reversal layer, and a Hadamard product.These components implement the proposed discriminator-gate architecture.
- Discriminator Gate: The discriminator gate reweights source samples using an estimated target-to-source density ratio, so the ratio acts as a gating function in classification.A GAN discriminator estimates the ratio from source-versus-target classification of input–label pairs, with implementation using D(F(x), y).
- Adversarial Objectives: The approach combines the gated classification objective with an augmented adversarial objective that matches joint distributions.The joint discriminator uses a dummy label to support both marginal and joint discrimination and can exploit unlabeled target data.
- Generality: The method is highly general: although presented through DANN, it can be applied to other adversarial transfer methods and extended to non-adversarial methods.Experiments adapt six deep methods from three categories to evaluate this generality.
- Evaluation: Experiments with six state-of-the-art deep methods show substantial performance improvements under potential negative-transfer conditions by largely avoiding negative transfer.The broader empirical study uses eight transfer learning methods and four benchmark datasets to examine the three identified factors.
5. Experiments
Experiments analyze negative transfer across controlled factors and benchmark datasets, then evaluate discriminator-gated variants that filter unrelated source data. The gated approach consistently improves deep transfer methods and reduces negative transfer, including under degenerate source conditions.
- Experimental Setup: Experiments use four benchmark datasets spanning small, moderate, and large domain shifts, including synthetic-to-real transfer.The datasets are Digits, Office-31, Office-Home, and VisDA.
- Experimental Setup: The study controls algorithm, divergence, and target factors to analyze when negative transfer occurs.Source perturbation rates control marginal and conditional divergence, while labeled target-data availability controls the target factor.
- Study of Negative Transfer: Increasing distribution divergence increases negative transfer, whereas similar domains can avoid it even under high input perturbation.DANN performance eventually falls below its target-only baseline as perturbation increases, while covariate-shift experiments on similar domains show no negative transfer.
- Study of Negative Transfer: Negative transfer depends on labeled target data: DANN is eventually outperformed by the target-only baseline as the labeled target ratio increases.Both DANN and the target-only baseline improve with more labeled target data, but DANN is more affected by divergence.
- Study of Negative Transfer: Transfer algorithms differ in vulnerability: KMM has a smaller negative transfer gap than TCA, while GTA is the most robust deep method in the comparison.The comparison uses the same training data under moderately different distributions and limited labeled target data.
- Evaluation of Discriminator Gate: Discriminator-gated models consistently improve all deep baselines and largely eliminate negative transfer, with larger gains on harder synthetic-to-real tasks.DANNgate achieves negative average NTG where DANN has positive NTG; under a degenerate source, the gated model reaches results comparable to the target-only baseline.
- Evaluation of Discriminator Gate: Ablations show that combining discriminator gating with joint-distribution estimation provides the full performance benefit.Gate-only and label-only variants outperform DANN but underperform DANNgate, while joint-distribution matching is important when marginal and conditional distributions shift.
- Evaluation of Discriminator Gate: Visualizations show the gate assigns lower weights to unrelated source samples, producing better target clustering after low-weight samples are removed.DANN assigns similar weights to perturbed and unperturbed source data, whereas DANNgate separates them through its weights.
6. Conclusion
The paper proposes a discriminator gate to avoid negative transfer and reports consistent improvements over base methods while largely avoiding negative transfer.
- The paper proposes a novel discriminator gate technique to avoid negative transfer.
- Negative transfer relates directly to the transfer algorithm, domain divergence, and target data.
- The method consistently improves base-method performance and largely avoids negative transfer.
A. Negative Transfer Definition
The paper argues that negative transfer cannot be defined independently of the transfer algorithm and identifies unresolved questions about rigorous measurement and causal factors.
- Prior work leaves the exact definition, test-time measurement, comparison baseline, and causes of negative transfer unresolved.
- The analysis concludes that negative transfer must be algorithm-specific.
- An algorithm-agnostic definition would be independent of the transfer algorithm being used.
A RPT (A(S, T )) > min
The paper rejects comparing a transfer algorithm with the best possible algorithm because that condition is impractical and can label source-ignoring algorithms as negative transfer.
- A source-ignoring algorithm can satisfy the proposed algorithm-agnostic condition despite performing no meaningful transfer.
- Comparing against the best possible algorithm is misleading because its risk cannot be evaluated at test time.
- The paper therefore defines negative transfer for a specific transfer algorithm using a tractable test-time condition.
- The resulting definition resolves whether negative transfer is algorithm-specific and what negative impact should be compared.