Asymmetric metric learning for knowledge transfer

Budnik, Mateusz; Avrithis, Yannis

doi:10.48550/arxiv.2006.16331

Cited by 8 publications

(25 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The reason for not seeing an improvement with the positive sampler might be that, in the task of compatible training, the most significant objective is to train the new model to map images to same features as the old model, rather than constraining the distance between positive pairs. Such a phenomenon has also been observed by the related work (Budnik & Avrithis, 2020). As shown in Table 3 of (Budnik & Avrithis, 2020), the method of "regression" (forces the old feature and new feature of the same image similar using cosine similarity) wins the asymmetric test (cross-model test), compared to other methods that regularize the positive pairs.…”

Section: A4 Ablation Studiessupporting

confidence: 60%

“…As shown in Table 6 of Appendix, similar results can be observed with or without a pre-sampler. The reason for the similar results might be that the most significant objective of compatible training is to train the new model to map images to same features as the old model rather than constraining the distance between positive pairs, according to (Budnik & Avrithis, 2020). Note that the negative pair is built by the new and old features of inter-class instances in the mini-batch, i.e., the intra-class samples would not be treated as the negatives in our instance discrimination-like compatible training.…”

Section: Revisit Conventional Compatible Trainingmentioning

confidence: 96%

“…Shen et al (2020) first formulated the problem of backward-compatible learning and proposed to utilize the old classifier for supervising the compatible feature learning. Budnik & Avrithis (2020) made an effort to reduce the gap between the large model and a small one with-out the parametric classifier. introduced a projection head to transform the old features to be compatible with the new ones.…”

Section: Related Workmentioning

confidence: 99%

“…Being cast as either the metric learning task (Budnik & Avrithis, 2020) or the transformation learning task , the core of existing compatible training methods is to encourage the new feature of an image to approach the old features of the intra-class images while keeping away from the old features of different classes. We treat the compatible training as an instance discriminationlike task (i.e., the positive pair is the new and old features of the same instance) for efficient training in large-scale retrieval tasks.…”

Section: Revisit Conventional Compatible Trainingmentioning

confidence: 99%

“…Mean Average Precision (mAP) is adopted here for evaluating the retrieval performance. Following the evaluation protocols in (Budnik & Avrithis, 2020;Weyand et al, 2020), we use mAP@100 for GLDv2-test, and mAP@10 for ROxford and RParis. Since mAP measures the overall accuracy with the consideration of both positive flips and negative flips, we define a metric, namely negative flip rate (NFR), to measure the degree of model regression in hot-refresh model upgrades.…”

Section: Implementation Detailsmentioning

confidence: 99%

See 4 more Smart Citations

Hot-Refresh Model Upgrades with Regression-Alleviating Compatible Training in Image Retrieval

Zhang¹,

Ge²,

Shen³

et al. 2022

Preprint

View full text Add to dashboard Cite

The task of hot-refresh model upgrades of image retrieval systems plays an essential role in the industry but has never been investigated in academia before. Conventional cold-refresh model upgrades can only deploy new models after the gallery is overall backfilled, taking weeks or even months for massive data. In contrast, hot-refresh model upgrades deploy the new model immediately and then gradually improve the retrieval accuracy by backfilling the gallery on-the-fly.Compatible training has made it possible, however, the problem of model regression * with negative flips poses a great challenge to the stable improvement of user experience. We argue that it is mainly due to the fact that new-to-old positive query-gallery pairs may show less similarity than new-to-new negative pairs. To solve the problem, we introduce a Regression-Alleviating Compatible Training (RACT) method to properly constrain the feature compatibility while reducing negative flips. The core is to encourage the new-to-old positive pairs to be more similar than both the new-to-old negative pairs and the new-to-new negative pairs. An efficient uncertainty-based backfilling strategy is further introduced to fasten accuracy improvements. Extensive experiments on large-scale retrieval benchmarks (e.g., Google Landmark) demonstrate that our RACT effectively alleviates the model regression for one more step towards seamless model upgrades § . † Corresponding authors. ‡ Work done when Binjie, Yantao and Yu are at ARC Lab, Tencent PCG. * Throughout this paper, the term model regression indicates the degradation of retrieval performance when the hot-refresh model upgrades proceed.§ The code will be available at https://github.com/binjiezhang/RACT_ICLR2022.

show abstract

Section: A4 Ablation Studiessupporting

confidence: 60%

Section: Revisit Conventional Compatible Trainingmentioning

confidence: 96%

Section: Related Workmentioning

confidence: 99%

Section: Revisit Conventional Compatible Trainingmentioning

confidence: 99%

Section: Implementation Detailsmentioning

confidence: 99%

See 3 more Smart Citations

Hot-Refresh Model Upgrades with Regression-Alleviating Compatible Training in Image Retrieval

Zhang¹,

Ge²,

Shen³

et al. 2022

Preprint

View full text Add to dashboard Cite

show abstract

Neighborhood Consensus Contrastive Learning for Backward-Compatible Representation

Chen

Lou

et al. 2022

AAAI

View full text Add to dashboard Cite

In object re-identification (ReID), the development of deep learning techniques often involves model updates and deployment. It is unbearable to re-embedding and re-index with the system suspended when deploying new models. Therefore, backward-compatible representation is proposed to enable ``new'' features to be compared with ``old'' features directly, which means that the database is active when there are both ``new'' and ``old'' features in it. Thus we can scroll-refresh the database or even do nothing on the database to update. The existing backward-compatible methods either require a strong overlap between old and new training data or simply conduct constraints at the instance level. Thus they are difficult in handling complicated cluster structures and are limited in eliminating the impact of outliers in old embeddings, resulting in a risk of damaging the discriminative capability of new features. In this work, we propose a Neighborhood Consensus Contrastive Learning (NCCL) method. With no assumptions about the new training data, we estimate the sub-cluster structures of old embeddings. A new embedding is constrained with multiple old embeddings in both embedding space and discrimination space at the sub-class level. The effect of outliers diminished, as the multiple samples serve as ``mean teachers''. Besides, we propose a scheme to filter the old embeddings with low credibility, further improving the compatibility robustness. Our method ensures the compatibility without impairing the accuracy of the new model. It can even improve the new model's accuracy in most scenarios.

show abstract

Darwinian Model Upgrades: Model Evolving with Selective Compatibility

Zhang

et al. 2023

AAAI

View full text Add to dashboard Cite

The traditional model upgrading paradigm for retrieval requires recomputing all gallery embeddings before deploying the new model (dubbed as "backfilling"), which is quite expensive and time-consuming considering billions of instances in industrial applications. BCT presents the first step towards backward-compatible model upgrades to get rid of backfilling. It is workable but leaves the new model in a dilemma between new feature discriminativeness and new-to-old compatibility due to the undifferentiated compatibility constraints. In this work, we propose Darwinian Model Upgrades (DMU), which disentangle the inheritance and variation in the model evolving with selective backward compatibility and forward adaptation, respectively. The old-to-new heritable knowledge is measured by old feature discriminativeness, and the gallery features, especially those of poor quality, are evolved in a lightweight manner to become more adaptive in the new latent space. We demonstrate the superiority of DMU through comprehensive experiments on large-scale landmark retrieval and face recognition benchmarks. DMU effectively alleviates the new-to-new degradation at the same time improving new-to-old compatibility, rendering a more proper model upgrading paradigm in large-scale retrieval systems.Code: https://github.com/TencentARC/OpenCompatible.

show abstract

Asymmetric metric learning for knowledge transfer

Cited by 8 publications

References 0 publications

Hot-Refresh Model Upgrades with Regression-Alleviating Compatible Training in Image Retrieval

Hot-Refresh Model Upgrades with Regression-Alleviating Compatible Training in Image Retrieval

Neighborhood Consensus Contrastive Learning for Backward-Compatible Representation

Darwinian Model Upgrades: Model Evolving with Selective Compatibility

Contact Info

Product

Resources

About