Characterizing and Avoiding Negative Transfer

Abstract

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.

Larger Norm More Transferable: An Adaptive Feature Norm Approach for Unsupervised Domain Adaptation

“…Though the notion of negative transfer has been well recognized within the DA community [30], its rigorous definition is still unclear [46]. A widely accepted description of negative transfer [30] is stated as transferring knowledge from the source can have a negative impact on the target learner.…”

“…While intuitive, how to evaluate it still remains open. Inspired by [46], we propose meaningful protocols to evaluate the robustness of a given algorithm especially under the more general partial setting. It is noteworthy that in this setting, the negative transfer is caused not only from the unrelated samples within the shared categories but also from the unrelated data from the source outlier classes.…”

Understanding the Effects of Pre-Training for Object Detectors via Eigenspectrum

“…With fine-tuning, learned parameters or features of source tasks may be forgotten after learning target tasks [29], and domain similarity between tasks being important for transfer learning [89]. Furthermore, transferring knowledge between dissimilar tasks may cause negative transfer [62,79]. Thus many works have discussed task difference between image classification and object detection [60,69,72,6] and [13] shows that object detectors I, which finetuned from ImageNet pre-trained models, can achieve high accuracy.…”

Integrating Machine Learning with Human Knowledge

“…Sometimes the performance may actually be worse than if the target task was trained alone. This is known as negative transfer and occurs when the source tasks are poorly suited for the target tasks ( Wang et al., 2019 ). It appears in human learning as well, e.g., learning to throw a baseball may be harder after learning to throw a football due to muscle memory making it difficult to adapt to a new throwing motion.…”

“…Currently, preventing negative transfer requires effective human intuition or experience. Research is conducted to develop methods that will quantitatively eliminate negative performance, including using a discriminator gate to assign different weights to each source task ( Wang et al., 2019 ) and using an iterative method that detects the source of the negative transfer to reduce class noise ( Gui et al., 2018 ).…”