UHKD: A Unified Framework for Heterogeneous Knowledge Distillation via Frequency-Domain Representations

The paper titled 'Preserving Cross-Modal Consistency for CLIP-based Class-Incremental Learning' addresses the challenges of class-incremental learning (CIL) in vision-language models like CLIP. It introduces a two-stage framework called DMC, which separates the adaptation of the vision encoder from the optimization of textual soft prompts. This approach aims to mitigate classifier bias and maintain cross-modal alignment, enhancing the model's ability to learn new categories without forgetting previously acquired knowledge.

The article introduces ERMoE, a new Mixture-of-Experts (MoE) architecture designed to enhance model capacity by addressing challenges in routing and expert specialization. ERMoE reparameterizes experts in an orthonormal eigenbasis and utilizes an 'Eigenbasis Score' for routing, which stabilizes expert utilization and improves interpretability. This approach aims to overcome issues of misalignment and load imbalances that have hindered previous MoE architectures.

The article introduces Autoregressive Representation Alignment (ARRA), a novel training framework designed to enhance text-to-image generation in autoregressive large language models (LLMs) without altering their architecture. ARRA achieves this by aligning the hidden states of LLMs with visual representations from external models through a global visual alignment loss and a hybrid token. Experimental results demonstrate that ARRA significantly reduces the Fréchet Inception Distance (FID) for models like LlamaGen, indicating improved coherence in generated images.

The paper titled 'Enhanced Structured Lasso Pruning with Class-wise Information' discusses advancements in neural network pruning methods. Traditional pruning techniques often overlook class-wise information, leading to potential loss of statistical data. This study introduces two new pruning schemes, sparse graph-structured lasso pruning with Information Bottleneck (sGLP-IB) and sparse tree-guided lasso pruning with Information Bottleneck (sTLP-IB), aimed at preserving statistical information while reducing model complexity.

The article discusses FedOrtho, a federated unlearning framework designed to enhance data unlearning in federated learning environments. It addresses the challenges of balancing forgetting and retention, particularly in non-IID settings. FedOrtho employs orthogonalized deep convolutional kernels and a one-shot soft pruning mechanism, achieving state-of-the-art performance on datasets like CIFAR-10 and TinyImageNet, with over 98% forgetting quality and 97% retention accuracy.

The paper introduces PrivDFS, a distributed feature-sharing framework designed for input-private inference in image classification. It addresses vulnerabilities in split inference that allow Data Reconstruction Attacks (DRAs) to reconstruct inputs with high fidelity. By fragmenting the intermediate representation and processing these fragments independently across a majority-honest set of servers, PrivDFS limits the reconstruction capability while maintaining accuracy within 1% of non-private methods.

The paper discusses advancements in out-of-distribution (OOD) detection, focusing on the integration of visual and textual modalities through large language models (LLMs). It introduces a method called Positive and Negative Prompt Supervision, which aims to improve OOD detection by using class-specific prompts that capture inter-class features. This approach addresses the limitations of negative prompts that may include non-ID features, potentially leading to suboptimal outcomes.

RiverScope is a newly developed high-resolution dataset aimed at improving the monitoring of rivers and surface water dynamics, which are crucial for understanding Earth's climate system. The dataset includes 1,145 high-resolution images covering 2,577 square kilometers, with expert-labeled river and surface water masks. This initiative addresses the challenges of monitoring narrow or sediment-rich rivers that are often inadequately represented in low-resolution satellite data.