Fine-tuning is arguably the most straightforward way to tailor a pre-trained model (e.g., a foundation model) to downstream applications, but it also comes with the risk of losing valuable knowledge the model had learned in pre-training. For example, fine-tuning a pre-trained classifier capable of recognizing a large number of classes to master a subset of classes at hand is shown to drastically degrade the model's accuracy in the other classes it had previously learned. As such, it is hard to further use the fine-tuned model when it encounters classes beyond the fine-tuning data. In this paper, we systematically dissect the issue, aiming to answer the fundamental question, "What has been damaged in the fine-tuned model?" To our surprise, we find that the fine-tuned model neither forgets the relationship among the other classes nor degrades the features to recognize these classes. Instead, the fine-tuned model often produces more discriminative features for these other classes, even if they were missing during fine-tuning! What really hurts the accuracy is the discrepant logit scales between the fine-tuning classes and the other classes, implying that a simple post-processing calibration would bring back the pre-trained model's capability and at the same time unveil the feature improvement over all classes. We conduct an extensive empirical study to demonstrate the robustness of our findings and provide preliminary explanations underlying them, suggesting new directions for future theoretical analysis.

Atrial fibrillation is a commonly encountered clinical arrhythmia associated with stroke and increased mortality. Since professional medical knowledge is required for annotation, exploiting a large corpus of ECGs to develop accurate supervised learning-based atrial fibrillation algorithms remains challenging. Self-supervised learning (SSL) is a promising recipe for generalized ECG representation learning, eliminating the dependence on expensive labeling. However, without well-designed incorporations of knowledge related to atrial fibrillation, existing SSL approaches typically suffer from unsatisfactory capture of robust ECG representations. In this paper, we propose an inter-intra period-aware ECG representation learning approach. Considering ECGs of atrial fibrillation patients exhibit the irregularity in RR intervals and the absence of P-waves, we develop specific pre-training tasks for interperiod and intraperiod representations, aiming to learn the single-period stable morphology representation while retaining crucial interperiod features. After further fine-tuning, our approach demonstrates remarkable AUC performances on the BTCH dataset, \textit{i.e.}, 0.953/0.996 for paroxysmal/persistent atrial fibrillation detection. On commonly used benchmarks of CinC2017 and CPSC2021, the generalization capability and effectiveness of our methodology are substantiated with competitive results.

Parameter-efficient transfer learning (PETL) has attracted significant attention lately, due to the increasing size of pre-trained models and the need to fine-tune them for superior downstream performance. This community-wide enthusiasm has sparked a plethora of approaches. Nevertheless, a systematic study to understand their performance and suitable application scenarios is lacking, leaving questions like "when to apply PETL" and "which approach to use" largely unanswered, especially in visual recognition. In this paper, we conduct a unifying empirical study of representative PETL approaches in the context of Vision Transformers (ViT). We systematically tune their hyper-parameters to fairly compare their accuracy on downstream tasks. Our study not only offers a valuable user guide but also unveils several new insights. First, if tuned carefully, different PETL approaches can obtain quite similar accuracy in the low-shot benchmark VTAB-1K. This includes simple approaches like fine-tuning the bias terms that were reported inferior. Second, though with similar accuracy, we find that PETL approaches make different mistakes and high-confidence predictions, likely due to their different inductive biases. Such an inconsistency (or complementariness) opens up the opportunity for ensemble methods, and we make preliminary attempts at this. Third, going beyond the commonly used low-shot tasks, we find that PETL is also useful in many-shot regimes -- it achieves comparable and sometimes better accuracy than full finetuning, using much fewer learnable parameters. Last but not least, we investigate PETL's ability to preserve a pre-trained model's robustness to distribution shifts (e.g., a CLIP backbone). Perhaps not surprisingly, PETL approaches outperform full fine-tuning alone. However, with weight-space ensembles, the fully fine-tuned model can better balance target (i.e., downstream) distribution and distribution shift performance, suggesting a future research direction for PETL Pre-training and then fine-tuning has become the standard practice to tackle visual recognition problems (Bommasani et al., 2021). The community-wide enthusiasm for open-sourcing has made it possible to access large, powerful pre-trained models learned from a gigantic amount of data, e.g., ImageNet-21K (Ridnik et al., 2021) or LAION-5B (Schuhmann et al., 2022). More research focus has thus been on how to fine-tune such large models (Yu et al., 2023a). Among existing efforts, parameter-efficient transfer learning (PETL), a.k.a parameter-efficient fine-tuning (PEFT), has attracted increasing attention lately (Han et al., 2024; Ding et al., 2023). Instead of fine-tuning the whole model (i.e., full fine-tuning) or the last fully connected layer (i.e., linear probing), PETL approaches seek to update or insert a relatively small number of parameters to the pre-trained model (Xin et al., 2024).

Semantic- and task-oriented communication has emerged as a promising approach to reducing the latency and bandwidth requirements of next-generation mobile networks by transmitting only the most relevant information needed to complete a specific task at the receiver. This is particularly advantageous for machine-oriented communication of high data rate content, such as images and videos, where the goal is rapid and accurate inference, rather than perfect signal reconstruction. While semantic- and task-oriented compression can be implemented in conventional communication systems, joint source-channel coding (JSCC) offers an alternative end-to-end approach by optimizing compression and channel coding together, or even directly mapping the source signal to the modulated waveform. Although all digital communication systems today rely on separation, thanks to its modularity, JSCC is known to achieve higher performance in finite blocklength scenarios, and to avoid cliff and the levelling-off effects in time-varying channel scenarios. This article provides an overview of the information theoretic foundations of JSCC, surveys practical JSCC designs over the decades, and discusses the reasons for their limited adoption in practical systems. We then examine the recent resurgence of JSCC, driven by the integration of deep learning techniques, particularly through DeepJSCC, highlighting its many surprising advantages in various scenarios. Finally, we discuss why it may be time to reconsider today's strictly separate architectures, and reintroduce JSCC to enable high-fidelity, low-latency communications in critical applications such as autonomous driving, drone surveillance, or wearable systems.

Lightweight and efficient neural network models for deep joint source-channel coding (JSCC) are crucial for semantic communications. In this paper, we propose a novel JSCC architecture, named MambaJSCC, that achieves state-of-the-art performance with low computational and parameter overhead. MambaJSCC utilizes the visual state space model with channel adaptation (VSSM-CA) blocks as its backbone for transmitting images over wireless channels, where the VSSM-CA primarily consists of the generalized state space models (GSSM) and the zero-parameter, zero-computational channel adaptation method (CSI-ReST). We design the GSSM module, leveraging reversible matrix transformations to express generalized scan expanding operations, and theoretically prove that two GSSM modules can effectively capture global information. We discover that GSSM inherently possesses the ability to adapt to channels, a form of endogenous intelligence. Based on this, we design the CSI-ReST method, which injects channel state information (CSI) into the initial state of GSSM to utilize its native response, and into the residual state to mitigate CSI forgetting, enabling effective channel adaptation without introducing additional computational and parameter overhead. Experimental results show that MambaJSCC not only outperforms existing JSCC methods (e.g., SwinJSCC) across various scenarios but also significantly reduces parameter size, computational overhead, and inference delay.

Semantic communication, as a revolutionary communication architecture, is considered a promising novel communication paradigm. Unlike traditional symbol-based error-free communication systems, semantic-based visual communication systems extract, compress, transmit, and reconstruct images at the semantic level. However, widely used image similarity evaluation metrics, whether pixel-based MSE or PSNR or structure-based MS-SSIM, struggle to accurately measure the loss of semantic-level information of the source during system transmission. This presents challenges in evaluating the performance of visual semantic communication systems, especially when comparing them with traditional communication systems. To address this, we propose a semantic evaluation metric -- SeSS (Semantic Similarity Score), based on Scene Graph Generation and graph matching, which shifts the similarity scores between images into semantic-level graph matching scores. Meanwhile, semantic similarity scores for tens of thousands of image pairs are manually annotated to fine-tune the hyperparameters in the graph matching algorithm, aligning the metric more closely with human semantic perception. The performance of the SeSS is tested on different datasets, including (1)images transmitted by traditional and semantic communication systems at different compression rates, (2)images transmitted by traditional and semantic communication systems at different signal-to-noise ratios, (3)images generated by large-scale model with different noise levels introduced, and (4)cases of images subjected to certain special transformations. The experiments demonstrate the effectiveness of SeSS, indicating that the metric can measure the semantic-level differences in semantic-level information of images and can be used for evaluation in visual semantic communication systems.

Large Vision Language Models (LVLMs) have recently achieved superior performance in various tasks on natural image and text data, which inspires a large amount of studies for LVLMs fine-tuning and training. Despite their advancements, there has been scant research on the robustness of these models against hallucination when fine-tuned on smaller datasets. In this study, we introduce a new benchmark dataset, the Medical Visual Hallucination Test (MedVH), to evaluate the hallucination of domain-specific LVLMs. MedVH comprises five tasks to evaluate hallucinations in LVLMs within the medical context, which includes tasks for comprehensive understanding of textual and visual input, as well as long textual response generation. Our extensive experiments with both general and medical LVLMs reveal that, although medical LVLMs demonstrate promising performance on standard medical tasks, they are particularly susceptible to hallucinations, often more so than the general models, raising significant concerns about the reliability of these domain-specific models. For medical LVLMs to be truly valuable in real-world applications, they must not only accurately integrate medical knowledge but also maintain robust reasoning abilities to prevent hallucination. Our work paves the way for future evaluations of these studies.

Precise perception of articulated objects is vital for empowering service robots. Recent studies mainly focus on point cloud, a single-modal approach, often neglecting vital texture and lighting details and assuming ideal conditions like optimal viewpoints, unrepresentative of real-world scenarios. To address these limitations, we introduce MARS, a novel framework for articulated object characterization. It features a multi-modal fusion module utilizing multi-scale RGB features to enhance point cloud features, coupled with reinforcement learning-based active sensing for autonomous optimization of observation viewpoints. In experiments conducted with various articulated object instances from the PartNet-Mobility dataset, our method outperformed current state-of-the-art methods in joint parameter estimation accuracy. Additionally, through active sensing, MARS further reduces errors, demonstrating enhanced efficiency in handling suboptimal viewpoints. Furthermore, our method effectively generalizes to real-world articulated objects, enhancing robot interactions. Code is available at https://github.com/robhlzeng/MARS.

In the field of 2D image generation modeling and representation learning, Masked Generative Encoder (MAGE) has demonstrated the synergistic potential between generative modeling and representation learning. Inspired by this, we propose Point-MAGE to extend this concept to point cloud data. Specifically, this framework first utilizes a Vector Quantized Variational Autoencoder (VQVAE) to reconstruct a neural field representation of 3D shapes, thereby learning discrete semantic features of point patches. Subsequently, by combining the masking model with variable masking ratios, we achieve synchronous training for both generation and representation learning. Furthermore, our framework seamlessly integrates with existing point cloud self-supervised learning (SSL) models, thereby enhancing their performance. We extensively evaluate the representation learning and generation capabilities of Point-MAGE. In shape classification tasks, Point-MAGE achieved an accuracy of 94.2% on the ModelNet40 dataset and 92.9% (+1.3%) on the ScanObjectNN dataset. Additionally, it achieved new state-of-the-art performance in few-shot learning and part segmentation tasks. Experimental results also confirmed that Point-MAGE can generate detailed and high-quality 3D shapes in both unconditional and conditional settings.

In this letter, we design a federated contrastive learning (FedCL) framework aimed at supporting personalized semantic communication. Our FedCL enables collaborative training of local semantic encoders across multiple clients and a global semantic decoder owned by the base station. This framework supports heterogeneous semantic encoders since it does not require client-side model aggregation. Furthermore, to tackle the semantic imbalance issue arising from heterogeneous datasets across distributed clients, we employ contrastive learning to train a semantic centroid generator (SCG). This generator obtains representative global semantic centroids that exhibit intra-semantic compactness and inter-semantic separability. Consequently, it provides superior supervision for learning discriminative local semantic features. Additionally, we conduct theoretical analysis to quantify the convergence performance of FedCL. Simulation results verify the superiority of the proposed FedCL framework compared to other distributed learning benchmarks in terms of task performance and robustness under different numbers of clients and channel conditions, especially in low signal-to-noise ratio and highly heterogeneous data scenarios.