D4C: Data-free Quantization for Contrastive Language-Image Pre-training Models

This study investigates the effectiveness of self-distillation (SD) in improving model performance using hyperparameter-tuned multi-stage SD with a linear classifier for binary classification on noisy Gaussian mixture data. The research employs statistical physics methods and finds that denoising through hard pseudo-labels significantly enhances SD performance, particularly in moderately sized datasets. Two heuristics are proposed to improve SD: early stopping and bias parameter fixing.

The article discusses the limitations of the CLIP model in capturing fine-grained details in remote sensing (RS) data. It highlights two main issues: the underutilization of object-level supervision in RS image-text datasets and the performance degradation of region-text alignment methods when applied to RS data. To address these challenges, the authors introduce the MGRS-200k dataset, which provides rich object-level textual supervision for improved RS region-category alignment.

Feature-map knowledge distillation (KD) is effective for convolutional networks but often fails for Vision Transformers (ViTs). A two-view representation analysis reveals that final-layer representations in ViTs are globally low-rank, suggesting that a compact student model should suffice for feature alignment. However, a token-level Spectral Energy Pattern analysis shows that individual tokens distribute energy across many channels, indicating a mismatch in encoding.

The article discusses a distributed learning problem where agents minimize a global objective function through information exchange over a network. The proposed method reduces communication by triggering it only when necessary and is agnostic to data distribution among agents, ensuring convergence even with distinct local data distributions. The convergence rate is analyzed in both convex and nonconvex settings, with numerical results demonstrating significant communication savings in distributed learning tasks on MNIST and CIFAR-10 datasets.

The paper presents an Unbiased Semantic Decoding (USD) strategy integrated with the Segment Anything Model (SAM) for few-shot segmentation tasks. This approach aims to enhance the model's generalization ability by extracting target information from both support and query sets simultaneously, addressing the limitations of previous methods that relied heavily on explicit prompts. The study highlights the potential of USD in improving segmentation accuracy across unknown classes.

The study explores the development of geometric and numerical concepts in humans through computer vision (CV) modeling. It builds on previous research indicating that CV models, despite being trained for image classification, can learn representations of geometric and numerical concepts akin to those of adults. The research specifically examines the ResNet-50 model, demonstrating that its performance aligns with developmental progressions observed in children.

The paper titled 'HV-Attack: Hierarchical Visual Attack for Multimodal Retrieval Augmented Generation' discusses a new method to exploit vulnerabilities in multimodal Retrieval-Augmented Generation (MRAG) systems. It highlights how imperceptible perturbations to image inputs can misalign and disrupt the generation process, posing significant safety concerns for Large Multimodal Models (LMMs). This research addresses the challenge of robustness in MRAG systems against such visual attacks.

The paper presents HASTEN (Hierarchical Semantic Tree Anchoring), a novel approach for Class-Incremental Learning (CIL) that integrates hierarchical information to mitigate catastrophic forgetting. It leverages external knowledge graphs to enhance the learning of visual and textual features, addressing the limitations of existing CLIP-based CIL methods that fail to capture inherent hierarchies in visual and linguistic concepts.