Text-to-image (T2I) generative models have recently emerged as a powerful tool, enabling the creation of photo-realistic images and giving rise to a multitude of applications. However, the effective integration of T2I models into fundamental image classification tasks remains an open question. A prevalent strategy to bolster image classification performance is through augmenting the training set with synthetic images generated by T2I models. In this study, we scrutinize the shortcomings of both current generative and conventional data augmentation techniques. Our analysis reveals that these methods struggle to produce images that are both faithful (in terms of foreground objects) and diverse (in terms of background contexts) for domain-specific concepts. To tackle this challenge, we introduce an innovative inter-class data augmentation method known as Diff-Mix (https://github.com/Zhicaiwww/Diff-Mix), which enriches the dataset by performing image translations between classes. Our empirical results demonstrate that Diff-Mix achieves a better balance between faithfulness and diversity, leading to a marked improvement in performance across diverse image classification scenarios, including few-shot, conventional, and long-tail classifications for domain-specific datasets.
This paper revives Densely Connected Convolutional Networks (DenseNets) and reveals the underrated effectiveness over predominant ResNet-style architectures. We believe DenseNets' potential was overlooked due to untouched training methods and traditional design elements not fully revealing their capabilities. Our pilot study shows dense connections through concatenation are strong, demonstrating that DenseNets can be revitalized to compete with modern architectures. We methodically refine suboptimal components - architectural adjustments, block redesign, and improved training recipes towards widening DenseNets and boosting memory efficiency while keeping concatenation shortcuts. Our models, employing simple architectural elements, ultimately surpass Swin Transformer, ConvNeXt, and DeiT-III - key architectures in the residual learning lineage. Furthermore, our models exhibit near state-of-the-art performance on ImageNet-1K, competing with the very recent models and downstream tasks, ADE20k semantic segmentation, and COCO object detection/instance segmentation. Finally, we provide empirical analyses that uncover the merits of the concatenation over additive shortcuts, steering a renewed preference towards DenseNet-style designs. Our code is available at https://github.com/naver-ai/rdnet.
The pursuit of learning robust representations without human supervision is a longstanding challenge. The recent advancements in self-supervised contrastive learning approaches have demonstrated high performance across various representation learning challenges. However, current methods depend on the random transformation of training examples, resulting in some cases of unrepresentative positive pairs that can have a large impact on learning. This limitation not only impedes the convergence of the learning process but the robustness of the learnt representation as well as requiring larger batch sizes to improve robustness to such bad batches. This paper attempts to alleviate the influence of false positive and false negative pairs by employing pairwise similarity calculations through the Fr\'echet ResNet Distance (FRD), thereby obtaining robust representations from unlabelled data. The effectiveness of the proposed method is substantiated by empirical results, where a linear classifier trained on self-supervised contrastive representations achieved an impressive 87.74\% top-1 accuracy on STL10 and 99.31\% on the Flower102 dataset. These results emphasize the potential of the proposed approach in pushing the boundaries of the state-of-the-art in self-supervised contrastive learning, particularly for image classification tasks.
Parameter-efficient fine-tuning for pre-trained Vision Transformers aims to adeptly tailor a model to downstream tasks by learning a minimal set of new adaptation parameters while preserving the frozen majority of pre-trained parameters. Striking a balance between retaining the generalizable representation capacity of the pre-trained model and acquiring task-specific features poses a key challenge. Currently, there is a lack of focus on guiding this delicate trade-off. In this study, we approach the problem from the perspective of Singular Value Decomposition (SVD) of pre-trained parameter matrices, providing insights into the tuning dynamics of existing methods. Building upon this understanding, we propose a Residual-based Low-Rank Rescaling (RLRR) fine-tuning strategy. This strategy not only enhances flexibility in parameter tuning but also ensures that new parameters do not deviate excessively from the pre-trained model through a residual design. Extensive experiments demonstrate that our method achieves competitive performance across various downstream image classification tasks, all while maintaining comparable new parameters. We believe this work takes a step forward in offering a unified perspective for interpreting existing methods and serves as motivation for the development of new approaches that move closer to effectively considering the crucial trade-off mentioned above. Our code is available at \href{https://github.com/zstarN70/RLRR.git}{https://github.com/zstarN70/RLRR.git}.
We present a generic Reinforcement Learning (RL) framework optimized for crafting adversarial attacks on different model types spanning from ECG signal analysis (1D), image classification (2D), and video classification (3D). The framework focuses on identifying sensitive regions and inducing misclassifications with minimal distortions and various distortion types. The novel RL method outperforms state-of-the-art methods for all three applications, proving its efficiency. Our RL approach produces superior localization masks, enhancing interpretability for image classification and ECG analysis models. For applications such as ECG analysis, our platform highlights critical ECG segments for clinicians while ensuring resilience against prevalent distortions. This comprehensive tool aims to bolster both resilience with adversarial training and transparency across varied applications and data types.
Remote sensing image classification forms the foundation of various understanding tasks, serving a crucial function in remote sensing image interpretation. The recent advancements of Convolutional Neural Networks (CNNs) and Transformers have markedly enhanced classification accuracy. Nonetheless, remote sensing scene classification remains a significant challenge, especially given the complexity and diversity of remote sensing scenarios and the variability of spatiotemporal resolutions. The capacity for whole-image understanding can provide more precise semantic cues for scene discrimination. In this paper, we introduce RSMamba, a novel architecture for remote sensing image classification. RSMamba is based on the State Space Model (SSM) and incorporates an efficient, hardware-aware design known as the Mamba. It integrates the advantages of both a global receptive field and linear modeling complexity. To overcome the limitation of the vanilla Mamba, which can only model causal sequences and is not adaptable to two-dimensional image data, we propose a dynamic multi-path activation mechanism to augment Mamba's capacity to model non-causal data. Notably, RSMamba maintains the inherent modeling mechanism of the vanilla Mamba, yet exhibits superior performance across multiple remote sensing image classification datasets. This indicates that RSMamba holds significant potential to function as the backbone of future visual foundation models. The code will be available at \url{https://github.com/KyanChen/RSMamba}.
Adversarial attacks have demonstrated the vulnerability of Machine Learning (ML) image classifiers in Synthetic Aperture Radar (SAR) Automatic Target Recognition (ATR) systems. An adversarial attack can deceive the classifier into making incorrect predictions by perturbing the input SAR images, for example, with a few scatterers attached to the on-ground objects. Therefore, it is critical to develop robust SAR ATR systems that can detect potential adversarial attacks by leveraging the inherent uncertainty in ML classifiers, thereby effectively alerting human decision-makers. In this paper, we propose a novel uncertainty-aware SAR ATR for detecting adversarial attacks. Specifically, we leverage the capability of Bayesian Neural Networks (BNNs) in performing image classification with quantified epistemic uncertainty to measure the confidence for each input SAR image. By evaluating the uncertainty, our method alerts when the input SAR image is likely to be adversarially generated. Simultaneously, we also generate visual explanations that reveal the specific regions in the SAR image where the adversarial scatterers are likely to to be present, thus aiding human decision-making with hints of evidence of adversarial attacks. Experiments on the MSTAR dataset demonstrate that our approach can identify over 80% adversarial SAR images with fewer than 20% false alarms, and our visual explanations can identify up to over 90% of scatterers in an adversarial SAR image.
Convolutional Neural Networks (CNNs) have advanced significantly in visual representation learning and recognition. However, they face notable challenges in performance and computational efficiency when dealing with real-world, multi-scale image inputs. Conventional methods rescale all input images into a fixed size, wherein a larger fixed size favors performance but rescaling small size images to a larger size incurs digitization noise and increased computation cost. In this work, we carry out a comprehensive, layer-wise investigation of CNN models in response to scale variation, based on Centered Kernel Alignment (CKA) analysis. The observations reveal lower layers are more sensitive to input image scale variations than high-level layers. Inspired by this insight, we propose Multi-scale Unified Network (MUSN) consisting of multi-scale subnets, a unified network, and scale-invariant constraint. Our method divides the shallow layers into multi-scale subnets to enable feature extraction from multi-scale inputs, and the low-level features are unified in deep layers for extracting high-level semantic features. A scale-invariant constraint is posed to maintain feature consistency across different scales. Extensive experiments on ImageNet and other scale-diverse datasets, demonstrate that MSUN achieves significant improvements in both model performance and computational efficiency. Particularly, MSUN yields an accuracy increase up to 44.53% and diminishes FLOPs by 7.01-16.13% in multi-scale scenarios.
Resource efficiency plays an important role for machine learning nowadays. The energy and decision latency are two critical aspects to ensure a sustainable and practical application. Unfortunately, the energy consumption and decision latency are not robust against adversaries. Researchers have recently demonstrated that attackers can compute and submit so-called sponge examples at inference time to increase the energy consumption and decision latency of neural networks. In computer vision, the proposed strategy crafts inputs with less activation sparsity which could otherwise be used to accelerate the computation. In this paper, we analyze the mechanism how these energy-latency attacks reduce activation sparsity. In particular, we find that input uniformity is a key enabler. A uniform image, that is, an image with mostly flat, uniformly colored surfaces, triggers more activations due to a specific interplay of convolution, batch normalization, and ReLU activation. Based on these insights, we propose two new simple, yet effective strategies for crafting sponge examples: sampling images from a probability distribution and identifying dense, yet inconspicuous inputs in natural datasets. We empirically examine our findings in a comprehensive evaluation with multiple image classification models and show that our attack achieves the same sparsity effect as prior sponge-example methods, but at a fraction of computation effort. We also show that our sponge examples transfer between different neural networks. Finally, we discuss applications of our findings for the good by improving efficiency by increasing sparsity.
In this work, we introduce Mini-Gemini, a simple and effective framework enhancing multi-modality Vision Language Models (VLMs). Despite the advancements in VLMs facilitating basic visual dialog and reasoning, a performance gap persists compared to advanced models like GPT-4 and Gemini. We try to narrow the gap by mining the potential of VLMs for better performance and any-to-any workflow from three aspects, i.e., high-resolution visual tokens, high-quality data, and VLM-guided generation. To enhance visual tokens, we propose to utilize an additional visual encoder for high-resolution refinement without increasing the visual token count. We further construct a high-quality dataset that promotes precise image comprehension and reasoning-based generation, expanding the operational scope of current VLMs. In general, Mini-Gemini further mines the potential of VLMs and empowers current frameworks with image understanding, reasoning, and generation simultaneously. Mini-Gemini supports a series of dense and MoE Large Language Models (LLMs) from 2B to 34B. It is demonstrated to achieve leading performance in several zero-shot benchmarks and even surpasses the developed private models. Code and models are available at https://github.com/dvlab-research/MiniGemini.
Contrastive learning has demonstrated great effectiveness in representation learning especially for image classification tasks. However, there is still a shortage in the studies targeting regression tasks, and more specifically applications on hyperspectral data. In this paper, we propose a contrastive learning framework for the regression tasks for hyperspectral data. To this end, we provide a collection of transformations relevant for augmenting hyperspectral data, and investigate contrastive learning for regression. Experiments on synthetic and real hyperspectral datasets show that the proposed framework and transformations significantly improve the performance of regression models, achieving better scores than other state-of-the-art transformations.
Graph neural networks (GNNs) have become increasingly popular for classification tasks on graph-structured data. Yet, the interplay between graph topology and feature evolution in GNNs is not well understood. In this paper, we focus on node-wise classification, illustrated with community detection on stochastic block model graphs, and explore the feature evolution through the lens of the "Neural Collapse" (NC) phenomenon. When training instance-wise deep classifiers (e.g. for image classification) beyond the zero training error point, NC demonstrates a reduction in the deepest features' within-class variability and an increased alignment of their class means to certain symmetric structures. We start with an empirical study that shows that a decrease in within-class variability is also prevalent in the node-wise classification setting, however, not to the extent observed in the instance-wise case. Then, we theoretically study this distinction. Specifically, we show that even an "optimistic" mathematical model requires that the graphs obey a strict structural condition in order to possess a minimizer with exact collapse. Interestingly, this condition is viable also for heterophilic graphs and relates to recent empirical studies on settings with improved GNNs' generalization. Furthermore, by studying the gradient dynamics of the theoretical model, we provide reasoning for the partial collapse observed empirically. Finally, we present a study on the evolution of within- and between-class feature variability across layers of a well-trained GNN and contrast the behavior with spectral methods.
We introduce the $\textbf{O}$ne-shot $\textbf{P}$runing $\textbf{T}$echnique for $\textbf{I}$nterchangeable $\textbf{N}$etworks ($\textbf{OPTIN}$) framework as a tool to increase the efficiency of pre-trained transformer architectures $\textit{without requiring re-training}$. Recent works have explored improving transformer efficiency, however often incur computationally expensive re-training procedures or depend on architecture-specific characteristics, thus impeding practical wide-scale adoption. To address these shortcomings, the OPTIN framework leverages intermediate feature distillation, capturing the long-range dependencies of model parameters (coined $\textit{trajectory}$), to produce state-of-the-art results on natural language, image classification, transfer learning, and semantic segmentation tasks $\textit{without re-training}$. Given a FLOP constraint, the OPTIN framework will compress the network while maintaining competitive accuracy performance and improved throughput. Particularly, we show a $\leq 2$% accuracy degradation from NLP baselines and a $0.5$% improvement from state-of-the-art methods on image classification at competitive FLOPs reductions. We further demonstrate the generalization of tasks and architecture with comparative performance using Mask2Former for semantic segmentation and cnn-style networks. OPTIN presents one of the first one-shot efficient frameworks for compressing transformer architectures that generalizes well across different class domains, in particular: natural language and image-related tasks, without $\textit{re-training}$.
We present PlainMamba: a simple non-hierarchical state space model (SSM) designed for general visual recognition. The recent Mamba model has shown how SSMs can be highly competitive with other architectures on sequential data and initial attempts have been made to apply it to images. In this paper, we further adapt the selective scanning process of Mamba to the visual domain, enhancing its ability to learn features from two-dimensional images by (i) a continuous 2D scanning process that improves spatial continuity by ensuring adjacency of tokens in the scanning sequence, and (ii) direction-aware updating which enables the model to discern the spatial relations of tokens by encoding directional information. Our architecture is designed to be easy to use and easy to scale, formed by stacking identical PlainMamba blocks, resulting in a model with constant width throughout all layers. The architecture is further simplified by removing the need for special tokens. We evaluate PlainMamba on a variety of visual recognition tasks, achieving performance gains over previous non-hierarchical models and is competitive with hierarchical alternatives. For tasks requiring high-resolution inputs, in particular, PlainMamba requires much less computing while maintaining high performance. Code and models are available at: https://github.com/ChenhongyiYang/PlainMamba .
In the computer vision literature, many effective histogram-based features have been developed. These engineered features include local binary patterns and edge histogram descriptors among others and they have been shown to be informative features for a variety of computer vision tasks. In this paper, we explore whether these features can be learned through histogram layers embedded in a neural network and, therefore, be leveraged within deep learning frameworks. By using histogram features, local statistics of the feature maps from the convolution neural networks can be used to better represent the data. We present neural versions of local binary pattern and edge histogram descriptors that jointly improve the feature representation and perform image classification. Experiments are presented on benchmark and real-world datasets.
Modeling and visualizing relationships between tasks or datasets is an important step towards solving various meta-tasks such as dataset discovery, multi-tasking, and transfer learning. However, many relationships, such as containment and transferability, are naturally asymmetric and current approaches for representation and visualization (e.g., t-SNE) do not readily support this. We propose Task2Box, an approach to represent tasks using box embeddings -- axis-aligned hyperrectangles in low dimensional spaces -- that can capture asymmetric relationships between them through volumetric overlaps. We show that Task2Box accurately predicts unseen hierarchical relationships between nodes in ImageNet and iNaturalist datasets, as well as transferability between tasks in the Taskonomy benchmark. We also show that box embeddings estimated from task representations (e.g., CLIP, Task2Vec, or attribute based) can be used to predict relationships between unseen tasks more accurately than classifiers trained on the same representations, as well as handcrafted asymmetric distances (e.g., KL divergence). This suggests that low-dimensional box embeddings can effectively capture these task relationships and have the added advantage of being interpretable. We use the approach to visualize relationships among publicly available image classification datasets on popular dataset hosting platform called Hugging Face.
Background and objective: Employing deep learning models in critical domains such as medical imaging poses challenges associated with the limited availability of training data. We present a strategy for improving the performance and generalization capabilities of models trained in low-data regimes. Methods: The proposed method starts with a pre-training phase, where features learned in a self-supervised learning setting are disentangled to improve the robustness of the representations for downstream tasks. We then introduce a meta-fine-tuning step, leveraging related classes between meta-training and meta-testing phases but varying the granularity level. This approach aims to enhance the model's generalization capabilities by exposing it to more challenging classification tasks during meta-training and evaluating it on easier tasks but holding greater clinical relevance during meta-testing. We demonstrate the effectiveness of the proposed approach through a series of experiments exploring several backbones, as well as diverse pre-training and fine-tuning schemes, on two distinct medical tasks, i.e., classification of prostate cancer aggressiveness from MRI data and classification of breast cancer malignity from microscopic images. Results: Our results indicate that the proposed approach consistently yields superior performance w.r.t. ablation experiments, maintaining competitiveness even when a distribution shift between training and evaluation data occurs. Conclusion: Extensive experiments demonstrate the effectiveness and wide applicability of the proposed approach. We hope that this work will add another solution to the arsenal of addressing learning issues in data-scarce imaging domains.
This paper introduces TinySaver, an early-exit-like dynamic model compression approach which employs tiny models to substitute large models adaptively. Distinct from traditional compression techniques, dynamic methods like TinySaver can leverage the difficulty differences to allow certain inputs to complete their inference processes early, thereby conserving computational resources. Most existing early exit designs are implemented by attaching additional network branches to the model's backbone. Our study, however, reveals that completely independent tiny models can replace a substantial portion of the larger models' job with minimal impact on performance. Employing them as the first exit can remarkably enhance computational efficiency. By searching and employing the most appropriate tiny model as the computational saver for a given large model, the proposed approaches work as a novel and generic method to model compression. This finding will help the research community in exploring new compression methods to address the escalating computational demands posed by rapidly evolving AI models. Our evaluation of this approach in ImageNet-1k classification demonstrates its potential to reduce the number of compute operations by up to 90\%, with only negligible losses in performance, across various modern vision models.
Deep learning models face persistent challenges in training, particularly due to internal covariate shift and label shift. While single-mode normalization methods like Batch Normalization partially address these issues, they are constrained by batch size dependencies and limiting distributional assumptions. Multi-mode normalization techniques mitigate these limitations but struggle with computational demands when handling diverse Gaussian distributions. In this paper, we introduce a new approach to multi-mode normalization that leverages prior knowledge to improve neural network representations. Our method organizes data into predefined structures, or "contexts", prior to training and normalizes based on these contexts, with two variants: Context Normalization (CN) and Context Normalization - Extended (CN-X). When contexts are unavailable, we introduce Adaptive Context Normalization (ACN), which dynamically builds contexts in the latent space during training. Across tasks in image classification, domain adaptation, and image generation, our methods demonstrate superior convergence and performance.
Deep learning has become an extremely powerful tool for complex tasks such as image classification and segmentation. The medical industry often lacks high-quality, balanced datasets, which can be a challenge for deep learning algorithms that need sufficiently large amounts of data to train and increase their performance. This is especially important in the context of kidney issues such as for stones, cysts and tumors. We used deep learning models for this study to classify or detect several types of kidney diseases. We use different classification models, such as VGG-19, (CNNs) Convolutional Neural Networks, ResNet-101, VGG-16, ResNet-50, and DenseNet-169, which can be enhanced through techniques such as classification, segmentation, and transfer learning. These algorithms can help improve model accuracy by allowing them to learn from multiple datasets. This technique has the potential to revolutionize the diagnosis and treatment of kidney problems as it enables more accurate and effective classification of CT-scan images. This may ultimately lead to better patient outcomes and improved overall health outcomes.