Unsupervised Visual Representation Learning by Online Constrained K-Means

Qian, Quan; Xu, Yuanhong; Hu, Juhua; Li, Hao; Jin, Rong

doi:10.1109/cvpr52688.2022.01614

Cited by 16 publications

(5 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To eliminate the cost of labeling, self-supervised learning is developed to obtain pre-trained model from unlabeled data. Many pretext tasks are proposed for effective learning, e.g., instance discrimination that considers each instance as an individual class and optimizes random augmentations from the same instance [6,18], cluster discrimination that explores relationship between different instances [5,32] and masked image modeling that leverages information within each image [16]. Moreover, [43] demonstrates that selfsupervised pre-training improves supervised pre-training with strong data augmentations.…”

Section: Visual Pre-trainingmentioning

confidence: 99%

“…Consequently, a lot of research efforts have been devoted to exploring unsupervised pre-training, which can eliminate the labeling cost for a vast amount of instances. Many effective self-supervised methods have been developed, e.g., instance discrimination [7,18], cluster discrimination [5,32], and masked image modeling [17]. Compared to the supervised counterparts, fine-tuning pre-trained models from these unsupervised methods can achieve comparable or even better performance on downstream tasks.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Beijing Language and Culture University, China

Wang¹

2022

The Experience of Examining the PhD

View full text Add to dashboard Cite

In the study of network synchronization, an outstanding question of both theoretical and practical significance is how to allocate a given set of heterogenous oscillators on a complex network in order for improving the synchronization performance. Whereas methods have been proposed to address this question in literature, the methods are based on accurate models describing the system dynamics, which, however, are normally unavailable in realistic situations. Here we show that this question can be addressed by the model-free technique of feed-forward neural network (FNN) in machine learning. Specifically, we measure the synchronization performance of a number of allocation schemes and use the measured data to train a machine. It is found that the trained machine is able to not only infer the synchronization performance of any new allocation scheme, but also find from a huge amount of candidates the optimal allocation scheme for synchronization.

show abstract

Section: Visual Pre-trainingmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Beijing Language and Culture University, China

Wang¹

2022

The Experience of Examining the PhD

View full text Add to dashboard Cite

show abstract

“…Unsupervised representation learning has the potential to overcome data scarcity and lack of availability by revealing previously hidden patterns and clusters. Clustering-based methods for this purpose may use either k-means [37]- [39] or Gaussian mixture models [40] to classify images with shared visual characteristics. The clusters may be used to create a group of visual prototypes that can serve as a feature representation for the subsequent tasks.…”

Section: Unsupervised Learningmentioning

confidence: 99%

VRL-IQA: Visual Representation Learning for Image Quality Assessment

Aslam,

Wei,

Ahmed

et al. 2024

IEEE Access

View full text Add to dashboard Cite

The growing adoption of digital multimedia devices and the greater reliance on compression and wireless channels for data transmission has brought renewed focus to the traditional challenge of evaluating image quality. Image Quality Assessment (IQA) is needed to optimize bit rate, compression, or processing and communication strategies for these multimedia technologies. Visual representation learning enables the model to undertake upstream training on large-scale data and then fine-tune the model on downstream data using fewer training samples. Data annotation for IQA is expensive due to the difficulty of grading a picture's quality, the need to gather quality labels from numerous observers, and the diversity of perceptual quality and content of the images. This challenge has limited the amount of the labeled training dataset for IQA to a few thousand. In this study, a deep Convolutional Neural Network is trained on a largescale image dataset produced by simulating 165 distortion scenarios on 150,000 images resulting in 24.75 million distorted images. These images are labeled via an ensemble of full-reference quality assessment models which assign the quality rating to each distorted image by using its reference image. This trained model is fine-tuned on two datasets TID2013 and Kadid-10K datasets containing simulated distortions and two datasets KonIQ-10K and BIQ2021 containing authentic distortions. The fine-tuning performance has resulted in state-of-the-art IQA performance and yielded a Spearman's correlation coefficient of 0.921, 0.893, 0.884, and 0.793, respectively. Moreover, comparison with the ImageNet pre-trained model revealed that the proposed VRL-IQA model provides higher performance in terms of Pearson and Spearman's correlations and achieves the validation criteria with fewer epochs than the ImageNet pre-trained model. These findings contribute to the advancement of IQA, offering a promising approach for robust and accurate quality prediction in various applications.

show abstract

“…Challenge 2: How to avoid "training collapse" in unsupervised dynamics disentanglement? Despite the great success in unsupervised representation learning [12,13,14], it remains a challenge to disentangle the controllable and noncontrollable dynamic patterns in non-stationary visual scenes. One potential solution is to employ modular structures that learn different dynamics in separate branches.…”

Section: Key Challengesmentioning

confidence: 99%

Model-Based Reinforcement Learning With Isolated Imaginations

Pan,

Zhu,

Zheng

et al. 2024

IEEE Trans. Pattern Anal. Mach. Intell.

View full text Add to dashboard Cite

World models learn the consequences of actions in vision-based interactive systems. However, in practical scenarios like autonomous driving, noncontrollable dynamics that are independent or sparsely dependent on action signals often exist, making it challenging to learn effective world models. To address this issue, we propose Iso-Dream++, a model-based reinforcement learning approach that has two main contributions. First, we optimize the inverse dynamics to encourage the world model to isolate controllable state transitions from the mixed spatiotemporal variations of the environment. Second, we perform policy optimization based on the decoupled latent imaginations, where we roll out noncontrollable states into the future and adaptively associate them with the current controllable state. This enables long-horizon visuomotor control tasks to benefit from isolating mixed dynamics sources in the wild, such as self-driving cars that can anticipate the movement of other vehicles, thereby avoiding potential risks. On top of our previous work [1], we further consider the sparse dependencies between controllable and noncontrollable states, address the training collapse problem of state decoupling, and validate our approach in transfer learning setups. Our empirical study demonstrates that Iso-Dream++ outperforms existing reinforcement learning models significantly on CARLA and DeepMind Control.

show abstract

Unsupervised Visual Representation Learning by Online Constrained K-Means

Cited by 16 publications

References 14 publications

Beijing Language and Culture University, China

Beijing Language and Culture University, China

VRL-IQA: Visual Representation Learning for Image Quality Assessment

Model-Based Reinforcement Learning With Isolated Imaginations

Contact Info

Product

Resources

About