MixMask: Enhancing Siamese Convolutional Networks with a Filling-Based Masking Strategy for Self-Supervised Learning
MixMask, a novel filling-based masking strategy for Siamese Convolutional Networks, improves self-supervised learning by replacing erased image regions with content from other images, thereby preserving global features crucial for contrastive learning.
Learning Robust and Predictive Neural Representations by Straightening Image Sequences for Improved Object Recognition
Training neural networks to learn representations that follow straight trajectories over time in response to sequences of transformed images leads to more robust and predictive models for object recognition compared to traditional invariance-based self-supervised learning methods.
C-JEPA: Enhancing Joint-Embedding Predictive Architecture with Contrastive Self-supervised Learning for Improved Visual Representation Learning
Integrating Variance-Invariance-Covariance Regularization (VICReg) into the Joint-Embedding Predictive Architecture (JEPA) significantly improves the stability and quality of visual representation learning by preventing model collapse and enhancing the learning of meaningful patch representations.
Pseudo-Label Refinement Using Hierarchical Clustering and Projection for Improved Self-Supervised Learning in Person Re-Identification
This paper introduces a novel pseudo-label refinement algorithm (SLR) that leverages cluster-label projection and hierarchical clustering to improve the accuracy of self-supervised learning systems, particularly in the context of person re-identification using unsupervised domain adaptation.
DINO 自監督方法系列中的部分原型坍塌問題
儘管現有的正則化技術，如居中和銳化，可以防止 DINO 自監督學習方法中出現完全表徵坍塌，但部分原型坍塌仍然是一個問題，導致原型存在顯著的冗餘，並阻礙了更細粒度、資訊更豐富的表徵學習。
DINO 계열 자기 지도 학습 방법에서 발생하는 부분 프로토타입 붕괴에 관하여
DINO 계열 자기 지도 학습 방법에서 발생하는 부분 프로토타입 붕괴 현상을 해결하여 프로토타입의 활용도를 높이고, 특히 Long-Tailed 데이터셋에서의 성능을 향상시키는 KoLeo-proto 정규화 기법을 제안합니다.
Mitigating Partial Prototype Collapse in Self-Supervised Learning: The Case of the DINO Family
The DINO family of self-supervised learning methods suffers from partial prototype collapse, where many prototypes become redundant, limiting the effectiveness of learned representations. This paper proposes a novel regularization technique, KoLeo-proto, to encourage diverse prototypes and improve performance on various downstream tasks, especially for long-tailed and fine-grained datasets.
A Probabilistic Model for Understanding and Improving Predictive Self-Supervised Learning
This paper introduces a generative probabilistic model, the SSL Model, to explain and unify various predictive self-supervised learning (SSL) methods, revealing their limitations in capturing style information and proposing a novel generative SSL approach, SimVAE, that outperforms existing methods in style retrieval tasks and achieves comparable or superior performance in content retrieval.
Conditional Augmentation-aware Self-supervised Learning with Conditioned Projector: Enhancing Sensitivity to Augmentations in Self-Supervised Representations
By conditioning the projector network with augmentation information, CASSLE enables self-supervised models to retain sensitivity to data augmentations, leading to improved performance on downstream tasks that rely on augmentation-affected features.
Failure-Proof Non-Contrastive Self-Supervised Learning with FALCON
This research paper introduces FALCON, a novel approach to non-contrastive self-supervised learning that guarantees the avoidance of common failure modes like representation, dimensional, cluster, and intracluster collapses, leading to improved generalization in downstream tasks.
© 2024 by Linnk AI