Bayesian neural networks (BNNs) treat neural network weights as random variables, which aim to provide posterior uncertainty estimates and avoid overfitting by performing inference on the posterior weights. However, the selection of appropriate prior distributions remains a challenging task, and BNNs may suffer from catastrophic inflated variance or poor predictive performance when poor choices are made for the priors. Existing BNN designs apply different priors to weights, while the behaviours of these priors make it difficult to sufficiently shrink noisy signals or they are prone to overshrinking important signals in the weights. To alleviate this problem, we propose a novel R2D2-Net, which imposes the R^2-induced Dirichlet Decomposition (R2D2) prior to the BNN weights. The R2D2-Net can effectively shrink irrelevant coefficients towards zero, while preventing key features from over-shrinkage. To approximate the posterior distribution of weights more accurately, we further propose a variational Gibbs inference algorithm that combines the Gibbs updating procedure and gradient-based optimization. This strategy enhances stability and consistency in estimation when the variational objective involving the shrinkage parameters is non-convex. We also analyze the evidence lower bound (ELBO) and the posterior concentration rates from a theoretical perspective. Experiments on both natural and medical image classification and uncertainty estimation tasks demonstrate satisfactory performance of our method.
This paper introduces Uniform Orthogonal Reinitialization Adaptation (UORA), a novel parameter-efficient fine-tuning (PEFT) approach for Large Language Models (LLMs). UORA achieves state-of-the-art performance and parameter efficiency by leveraging a low-rank approximation method to reduce the number of trainable parameters. Unlike existing methods such as LoRA and VeRA, UORA employs an interpolation-based reparametrization mechanism that selectively reinitializes rows and columns in frozen projection matrices, guided by the vector magnitude heuristic. This results in substantially fewer trainable parameters compared to LoRA and outperforms VeRA in computation and storage efficiency. Comprehensive experiments across various benchmarks demonstrate UORA's superiority in achieving competitive fine-tuning performance with negligible computational overhead. We demonstrate its performance on GLUE and E2E benchmarks and its effectiveness in instruction-tuning large language models and image classification models. Our contributions establish a new paradigm for scalable and resource-efficient fine-tuning of LLMs.
This paper focuses on using few-shot learning to improve the accuracy of classifying OCT diagnosis images with major and rare classes. We used the GAN-based augmentation strategy as a baseline and introduced several novel methods to further enhance our model. The proposed strategy contains U-GAT-IT for improving the generative part and uses the data balance technique to narrow down the skew of accuracy between all categories. The best model obtained was built with CBAM attention mechanism and fine-tuned InceptionV3, and achieved an overall accuracy of 97.85%, representing a significant improvement over the original baseline.
Using massive datasets, foundation models are large-scale, pre-trained models that perform a wide range of tasks. These models have shown consistently improved results with the introduction of new methods. It is crucial to analyze how these trends impact the medical field and determine whether these advancements can drive meaningful change. This study investigates the application of recent state-of-the-art foundation models, DINOv2, MAE, VMamba, CoCa, SAM2, and AIMv2, for medical image classification. We explore their effectiveness on datasets including CBIS-DDSM for mammography, ISIC2019 for skin lesions, APTOS2019 for diabetic retinopathy, and CHEXPERT for chest radiographs. By fine-tuning these models and evaluating their configurations, we aim to understand the potential of these advancements in medical image classification. The results indicate that these advanced models significantly enhance classification outcomes, demonstrating robust performance despite limited labeled data. Based on our results, AIMv2, DINOv2, and SAM2 models outperformed others, demonstrating that progress in natural domain training has positively impacted the medical domain and improved classification outcomes. Our code is publicly available at: https://github.com/sajjad-sh33/Medical-Transfer-Learning.
Multimodal Large Language Models (MLLMs) have demonstrated impressive capabilities across tasks, yet they often exhibit difficulty in distinguishing task-relevant from irrelevant signals, particularly in tasks like Visual Question Answering (VQA), which can lead to susceptibility to misleading or spurious inputs. We refer to this broader limitation as the Cross-Modality Competency Problem: the model's inability to fairly evaluate all modalities. This vulnerability becomes more evident in modality-specific tasks such as image classification or pure text question answering, where models are expected to rely solely on one modality. In such tasks, spurious information from irrelevant modalities often leads to significant performance degradation. We refer to this failure as Modality Interference, which serves as a concrete and measurable instance of the cross-modality competency problem. We further design a perturbation-based causal diagnostic experiment to verify and quantify this problem. To mitigate modality interference, we propose a novel framework to fine-tune MLLMs, including perturbation-based data augmentations with both heuristic perturbations and adversarial perturbations via Projected Gradient Descent (PGD), and a consistency regularization strategy applied to model outputs with original and perturbed inputs. Experiments on multiple benchmark datasets (image-heavy, text-heavy, and VQA tasks) and multiple model families with different scales demonstrate significant improvements in robustness and cross-modality competency, indicating our method's effectiveness in boosting unimodal reasoning ability while enhancing performance on multimodal tasks.
Medical image classification is critical for clinical decision-making, yet demands for accuracy, interpretability, and generalizability remain challenging. This paper introduces EVM-Fusion, an Explainable Vision Mamba architecture featuring a novel Neural Algorithmic Fusion (NAF) mechanism for multi-organ medical image classification. EVM-Fusion leverages a multipath design, where DenseNet and U-Net based pathways, enhanced by Vision Mamba (Vim) modules, operate in parallel with a traditional feature pathway. These diverse features are dynamically integrated via a two-stage fusion process: cross-modal attention followed by the iterative NAF block, which learns an adaptive fusion algorithm. Intrinsic explainability is embedded through path-specific spatial attention, Vim {\Delta}-value maps, traditional feature SE-attention, and cross-modal attention weights. Experiments on a diverse 9-class multi-organ medical image dataset demonstrate EVM-Fusion's strong classification performance, achieving 99.75% test accuracy and provide multi-faceted insights into its decision-making process, highlighting its potential for trustworthy AI in medical diagnostics.
In recent years, image classification, as a core task in computer vision, relies on high-quality labelled data, which restricts the wide application of deep learning models in practical scenarios. To alleviate the problem of insufficient labelled samples, semi-supervised learning has gradually become a research hotspot. In this paper, we construct a semi-supervised image classification model based on Generative Adversarial Networks (GANs), and through the introduction of the collaborative training mechanism of generators, discriminators and classifiers, we achieve the effective use of limited labelled data and a large amount of unlabelled data, improve the quality of image generation and classification accuracy, and provide an effective solution for the task of image recognition in complex environments.
Prompt learning has emerged as a powerful paradigm for adapting vision-language models such as CLIP to downstream tasks. However, existing methods often overfit to seen data, leading to significant performance degradation when generalizing to novel classes or unseen domains. To address this limitation, we propose DiSa, a Directional Saliency-Aware Prompt Learning framework that integrates two complementary regularization strategies to enhance generalization. First, our Cross-Interactive Regularization (CIR) fosters cross-modal alignment by enabling cooperative learning between prompted and frozen encoders. Within CIR, a saliency-aware masking strategy guides the image encoder to prioritize semantically critical image regions, reducing reliance on less informative patches. Second, we introduce a directional regularization strategy that aligns visual embeddings with class-wise prototype features in a directional manner to prioritize consistency in feature orientation over strict proximity. This approach ensures robust generalization by leveraging stable prototype directions derived from class-mean statistics. Extensive evaluations on 11 diverse image classification benchmarks demonstrate that DiSa consistently outperforms state-of-the-art prompt learning methods across various settings, including base-to-novel generalization, cross-dataset transfer, domain generalization, and few-shot learning.
To address the challenges posed by the large number of parameters in existing remote sensing image classification models, which hinder deployment on resource-constrained devices, this paper proposes a lightweight classification method based on knowledge distillation. Specifically, G-GhostNet is adopted as the backbone network, leveraging feature reuse to reduce redundant parameters and significantly improve inference efficiency. In addition, a decoupled knowledge distillation strategy is employed, which separates target and non-target classes to effectively enhance classification accuracy. Experimental results on the RSOD and AID datasets demonstrate that, compared with the high-parameter VGG-16 model, the proposed method achieves nearly equivalent Top-1 accuracy while reducing the number of parameters by 6.24 times. This approach strikes an excellent balance between model size and classification performance, offering an efficient solution for deployment on resource-limited devices.
Natural images exhibit label diversity (clean vs. noisy) in noisy-labeled image classification and prevalence diversity (abundant vs. sparse) in long-tailed image classification. Similarly, medical images in universal lesion detection (ULD) exhibit substantial variations in image quality, encompassing attributes such as clarity and label correctness. How to effectively leverage training images with diverse qualities becomes a problem in learning deep models. Conventional training mechanisms, such as self-paced curriculum learning (SCL) and online hard example mining (OHEM), relieve this problem by reweighting images with high loss values. Despite their success, these methods still confront two challenges: (i) the loss-based measure of sample hardness is imprecise, preventing optimum handling of different cases, and (ii) there exists under-utilization in SCL or over-utilization OHEM with the identified hard samples. To address these issues, this paper revisits the minibatch sampling (MBS), a technique widely used in deep network training but largely unexplored concerning the handling of diverse-quality training samples. We discover that the samples within a minibatch influence each other during training; thus, we propose a novel Mixed-order Minibatch Sampling (MoMBS) method to optimize the use of training samples with diverse qualities. MoMBS introduces a measure that takes both loss and uncertainty into account to surpass a sole reliance on loss and allows for a more refined categorization of high-loss samples by distinguishing them as either poorly labeled and under represented or well represented and overfitted. We prioritize under represented samples as the main gradient contributors in a minibatch and keep them from the negative influences of poorly labeled or overfitted samples with a mixed-order minibatch sampling design.
We introduce the Convolutional Low-Rank Adaptation (CoLoRA) method, designed explicitly to overcome the inefficiencies found in current CNN fine-tuning methods. CoLoRA can be seen as a natural extension of the convolutional architectures of the Low-Rank Adaptation (LoRA) technique. We demonstrate the capabilities of our method by developing and evaluating models using the widely adopted CNN backbone pre-trained on ImageNet. We observed that this strategy results in a stable and accurate coarse-tuning procedure. Moreover, this strategy is computationally efficient and significantly reduces the number of parameters required for fine-tuning compared to traditional methods. Furthermore, our method substantially improves the speed and stability of training. Our case study focuses on classifying retinal diseases from optical coherence tomography (OCT) images, specifically using the OCTMNIST dataset. Experimental results demonstrate that a CNN backbone fine-tuned with CoLoRA surpasses nearly 1\% in accuracy. Such a performance is comparable to the Vision Transformer, State-space discrete, and Kolmogorov-Arnold network models.
This paper presents a fascinating find: By training an auto-regressive LLM model on text tokens, the text model inherently develops internally an ability to understand images and audio, thereby developing the ability to see and hear just by reading. Popular audio and visual LLM models fine-tune text LLM models to give text output conditioned on images and audio embeddings. On the other hand, our architecture takes in patches of images, audio waveforms or tokens as input. It gives us the embeddings or category labels typical of a classification pipeline. We show the generality of text weights in aiding audio classification for datasets FSD-50K and GTZAN. Further, we show this working for image classification on CIFAR-10 and Fashion-MNIST, as well on image patches. This pushes the notion of text-LLMs learning powerful internal circuits that can be utilized by activating necessary connections for various applications rather than training models from scratch every single time.
Federated Learning (FL) is a decentralized machine learning paradigm that enables clients to collaboratively train models while preserving data privacy. However, the coexistence of model and data heterogeneity gives rise to inconsistent representations and divergent optimization dynamics across clients, ultimately hindering robust global performance. To transcend these challenges, we propose Mosaic, a novel data-free knowledge distillation framework tailored for heterogeneous distributed environments. Mosaic first trains local generative models to approximate each client's personalized distribution, enabling synthetic data generation that safeguards privacy through strict separation from real data. Subsequently, Mosaic forms a Mixture-of-Experts (MoE) from client models based on their specialized knowledge, and distills it into a global model using the generated data. To further enhance the MoE architecture, Mosaic integrates expert predictions via a lightweight meta model trained on a few representative prototypes. Extensive experiments on standard image classification benchmarks demonstrate that Mosaic consistently outperforms state-of-the-art approaches under both model and data heterogeneity. The source code has been published at https://github.com/Wings-Of-Disaster/Mosaic.
Fully decentralized training of machine learning models offers significant advantages in scalability, robustness, and fault tolerance. However, achieving differential privacy (DP) in such settings is challenging due to the absence of a central aggregator and varying trust assumptions among nodes. In this work, we present a novel privacy analysis of decentralized gossip-based averaging algorithms with additive node-level noise, both with and without secure summation over each node's direct neighbors. Our main contribution is a new analytical framework based on a linear systems formulation that accurately characterizes privacy leakage across these scenarios. This framework significantly improves upon prior analyses, for example, reducing the R\'enyi DP parameter growth from $O(T^2)$ to $O(T)$, where $T$ is the number of training rounds. We validate our analysis with numerical results demonstrating superior DP bounds compared to existing approaches. We further illustrate our analysis with a logistic regression experiment on MNIST image classification in a fully decentralized setting, demonstrating utility comparable to central aggregation methods.
Accurate classification of skin lesions from dermatoscopic images is essential for diagnosis and treatment of skin cancer. In this study, we investigate the utility of a dermatology-specific foundation model, PanDerm, in comparison with two Vision Transformer (ViT) architectures (ViT base and Swin Transformer V2 base) for the task of skin lesion classification. Using frozen features extracted from PanDerm, we apply non-linear probing with three different classifiers, namely, multi-layer perceptron (MLP), XGBoost, and TabNet. For the ViT-based models, we perform full fine-tuning to optimize classification performance. Our experiments on the HAM10000 and MSKCC datasets demonstrate that the PanDerm-based MLP model performs comparably to the fine-tuned Swin transformer model, while fusion of PanDerm and Swin Transformer predictions leads to further performance improvements. Future work will explore additional foundation models, fine-tuning strategies, and advanced fusion techniques.
Deep neural networks for image classification remain vulnerable to adversarial examples -- small, imperceptible perturbations that induce misclassifications. In black-box settings, where only the final prediction is accessible, crafting targeted attacks that aim to misclassify into a specific target class is particularly challenging due to narrow decision regions. Current state-of-the-art methods often exploit the geometric properties of the decision boundary separating a source image and a target image rather than incorporating information from the images themselves. In contrast, we propose Targeted Edge-informed Attack (TEA), a novel attack that utilizes edge information from the target image to carefully perturb it, thereby producing an adversarial image that is closer to the source image while still achieving the desired target classification. Our approach consistently outperforms current state-of-the-art methods across different models in low query settings (nearly 70\% fewer queries are used), a scenario especially relevant in real-world applications with limited queries and black-box access. Furthermore, by efficiently generating a suitable adversarial example, TEA provides an improved target initialization for established geometry-based attacks.
This study aims to address the growing challenge of distinguishing computer-generated imagery (CGI) from authentic digital images across three different color spaces; RGB, YCbCr, and HSV. Given the limitations of existing classification methods in handling the complexity and variability of CGI, this research proposes a Swin Transformer based model for accurate differentiation between natural and synthetic images. The proposed model leverages the Swin Transformer's hierarchical architecture to capture local and global features for distinguishing CGI from natural images. Its performance was assessed through intra- and inter-dataset testing across three datasets: CiFAKE, JSSSTU, and Columbia. The model was evaluated individually on each dataset (D1, D2, D3) and on the combined datasets (D1+D2+D3) to test its robustness and domain generalization. To address dataset imbalance, data augmentation techniques were applied. Additionally, t-SNE visualization was used to demonstrate the feature separability achieved by the Swin Transformer across the selected color spaces. The model's performance was tested across all color schemes, with the RGB color scheme yielding the highest accuracy for each dataset. As a result, RGB was selected for domain generalization analysis and compared with other CNN-based models, VGG-19 and ResNet-50. The comparative results demonstrate the proposed model's effectiveness in detecting CGI, highlighting its robustness and reliability in both intra-dataset and inter-dataset evaluations. The findings of this study highlight the Swin Transformer model's potential as an advanced tool for digital image forensics, particularly in distinguishing CGI from natural images. The model's strong performance indicates its capability for domain generalization, making it a valuable asset in scenarios requiring precise and reliable image classification.
Image classification benchmark datasets such as CIFAR, MNIST, and ImageNet serve as critical tools for model evaluation. However, despite the cleaning efforts, these datasets still suffer from pervasive noisy labels and often contain missing labels due to the co-existing image pattern where multiple classes appear in an image sample. This results in misleading model comparisons and unfair evaluations. Existing label cleaning methods focus primarily on noisy labels, but the issue of missing labels remains largely overlooked. Motivated by these challenges, we present a comprehensive framework named REVEAL, integrating state-of-the-art pre-trained vision-language models (e.g., LLaVA, BLIP, Janus, Qwen) with advanced machine/human label curation methods (e.g., Docta, Cleanlab, MTurk), to systematically address both noisy labels and missing label detection in widely-used image classification test sets. REVEAL detects potential noisy labels and omissions, aggregates predictions from various methods, and refines label accuracy through confidence-informed predictions and consensus-based filtering. Additionally, we provide a thorough analysis of state-of-the-art vision-language models and pre-trained image classifiers, highlighting their strengths and limitations within the context of dataset renovation by revealing 10 observations. Our method effectively reveals missing labels from public datasets and provides soft-labeled results with likelihoods. Through human verifications, REVEAL significantly improves the quality of 6 benchmark test sets, highly aligning to human judgments and enabling more accurate and meaningful comparisons in image classification.
Exploring the trustworthiness of deep learning models is crucial, especially in critical domains such as medical imaging decision support systems. Conformal prediction has emerged as a rigorous means of providing deep learning models with reliable uncertainty estimates and safety guarantees. However, conformal prediction results face challenges due to the backbone model's struggles in domain-shifted scenarios, such as variations in different sources. To aim this challenge, this paper proposes a novel framework termed Conformal Ensemble of Vision Transformers (CE-ViTs) designed to enhance image classification performance by prioritizing domain adaptation and model robustness, while accounting for uncertainty. The proposed method leverages an ensemble of vision transformer models in the backbone, trained on diverse datasets including HAM10000, Dermofit, and Skin Cancer ISIC datasets. This ensemble learning approach, calibrated through the combined mentioned datasets, aims to enhance domain adaptation through conformal learning. Experimental results underscore that the framework achieves a high coverage rate of 90.38\%, representing an improvement of 9.95\% compared to the HAM10000 model. This indicates a strong likelihood that the prediction set includes the true label compared to singular models. Ensemble learning in CE-ViTs significantly improves conformal prediction performance, increasing the average prediction set size for challenging misclassified samples from 1.86 to 3.075.
We introduce GradPCA, an Out-of-Distribution (OOD) detection method that exploits the low-rank structure of neural network gradients induced by Neural Tangent Kernel (NTK) alignment. GradPCA applies Principal Component Analysis (PCA) to gradient class-means, achieving more consistent performance than existing methods across standard image classification benchmarks. We provide a theoretical perspective on spectral OOD detection in neural networks to support GradPCA, highlighting feature-space properties that enable effective detection and naturally emerge from NTK alignment. Our analysis further reveals that feature quality -- particularly the use of pretrained versus non-pretrained representations -- plays a crucial role in determining which detectors will succeed. Extensive experiments validate the strong performance of GradPCA, and our theoretical framework offers guidance for designing more principled spectral OOD detectors.