In contrastive learning, two views of an original image generated by
different augmentations are considered as a positive pair whose similarity is
required to be high. Moreover, two views of two different images are considered
as a negative pair, and their similarity is encouraged to be low. Normally, a
single similarity measure given by a single projection head is used to evaluate
positive and negative sample pairs, respectively. However, due to the various
augmentation strategies and varying intra-sample similarity, augmented views
from the same image are often not similar. Moreover, due to inter-sample
similarity, augmented views of two different images may be more similar than
augmented views from the same image. As such, enforcing a high similarity for
positive pairs and a low similarity for negative pairs may not always be
achievable, and in the case of some pairs, forcing so may be detrimental to the
performance. To address this issue, we propose to use multiple projection
heads, each producing a separate set of features. Our loss function for
pre-training emerges from a solution to the maximum likelihood estimation over
head-wise posterior distributions of positive samples given observations. The
loss contains the similarity measure over positive and negative pairs, each
re-weighted by an individual adaptive temperature that is regularized to
prevent ill solutions. Our adaptive multi-head contrastive learning (AMCL) can
be applied to and experimentally improves several popular contrastive learning
methods such as SimCLR, MoCo and Barlow Twins. Such improvement is consistent
under various backbones and linear probing epoches and is more significant when
multiple augmentation methods are used.

对比学习中，为了实现正样本对高相似度要求和负样本对低相似度要求，提出了使用多个投影头的自适应多头对比学习（AMCL）方法，通过使用每个投影头生成不同特征集合，以及通过自适应温度进行加权和正则化，针对问题进行改进。

自适应多头对比学习

Adaptive Multi-head Contrastive Learning

Asymmetric appearance between positive pair effectively reduces the risk of
representation degradation in contrastive learning. However, there are still a
mass of appearance similarities between positive pair constructed by the
existing methods, which inhibits the further representation improvement. In
this paper, we propose a novel asymmetric patch sampling strategy for
contrastive learning, to further boost the appearance asymmetry for better
representations. Specifically, dual patch sampling strategies are applied to
the given image, to obtain asymmetric positive pairs. First, sparse patch
sampling is conducted to obtain the first view, which reduces spatial
redundancy of image and allows a more asymmetric view. Second, a selective
patch sampling is proposed to construct another view with large appearance
discrepancy relative to the first one. Due to the inappreciable appearance
similarity between positive pair, the trained model is encouraged to capture
the similarity on semantics, instead of low-level ones. Experimental results
demonstrate that our proposed method significantly outperforms the existing
self-supervised methods on both ImageNet-1K and CIFAR dataset, e.g., 2.5%
finetune accuracy improvement on CIFAR100. Furthermore, our method achieves
state-of-the-art performance on downstream tasks, object detection and instance
segmentation on COCO.Additionally, compared to other self-supervised methods,
our method is more efficient on both memory and computation during training.
The source code is available at this https URL

本文提出了一种新的不对称补丁采样策略，用于对比学习，以进一步增强外观不对称性以获得更好的表示，并在 COCO 数据集上实现了最先进的物体检测和实例分割性能。