Vision Transformer (ViT) has emerged as a prominent backbone for computer vision. For more efficient ViTs, recent works lessen the quadratic cost of the self-attention layer by pruning or fusing the redundant tokens. However, these works faced the speed-accuracy trade-off caused by the loss of information. Here, we argue that token fusion needs to consider diverse relations between tokens to minimize information loss. In this paper, we propose a Multi-criteria Token Fusion (MCTF), that gradually fuses the tokens based on multi-criteria (e.g., similarity, informativeness, and size of fused tokens). Further, we utilize the one-step-ahead attention, which is the improved approach to capture the informativeness of the tokens. By training the model equipped with MCTF using a token reduction consistency, we achieve the best speed-accuracy trade-off in the image classification (ImageNet1K). Experimental results prove that MCTF consistently surpasses the previous reduction methods with and without training. Specifically, DeiT-T and DeiT-S with MCTF reduce FLOPs by about 44% while improving the performance (+0.5%, and +0.3%) over the base model, respectively. We also demonstrate the applicability of MCTF in various Vision Transformers (e.g., T2T-ViT, LV-ViT), achieving at least 31% speedup without performance degradation. Code is available at https://github.com/mlvlab/MCTF.
Medical data often exhibits distribution shifts, which cause test-time performance degradation for deep learning models trained using standard supervised learning pipelines. This challenge is addressed in the field of Domain Generalization (DG) with the sub-field of Single Domain Generalization (SDG) being specifically interesting due to the privacy- or logistics-related issues often associated with medical data. Existing disentanglement-based SDG methods heavily rely on structural information embedded in segmentation masks, however classification labels do not provide such dense information. This work introduces a novel SDG method aimed at medical image classification that leverages channel-wise contrastive disentanglement. It is further enhanced with reconstruction-based style regularization to ensure extraction of distinct style and structure feature representations. We evaluate our method on the complex task of multicenter histopathology image classification, comparing it against state-of-the-art (SOTA) SDG baselines. Results demonstrate that our method surpasses the SOTA by a margin of 1% in average accuracy while also showing more stable performance. This study highlights the importance and challenges of exploring SDG frameworks in the context of the classification task. The code is publicly available at https://github.com/BioMedIA-MBZUAI/ConDiSR
Capsule Neural Networks (CapsNets) is a novel architecture that utilizes vector-wise representations formed by multiple neurons. Specifically, the Dynamic Routing CapsNets (DR-CapsNets) employ an affine matrix and dynamic routing mechanism to train capsules and acquire translation-equivariance properties, enhancing its robustness compared to traditional Convolutional Neural Networks (CNNs). Echocardiograms, which capture moving images of the heart, present unique challenges for traditional image classification methods. In this paper, we explore the potential of DR-CapsNets and propose CardioCaps, a novel attention-based DR-CapsNet architecture for class-imbalanced echocardiogram classification. CardioCaps comprises two key components: a weighted margin loss incorporating a regression auxiliary loss and an attention mechanism. First, the weighted margin loss prioritizes positive cases, supplemented by an auxiliary loss function based on the Ejection Fraction (EF) regression task, a crucial measure of cardiac function. This approach enhances the model's resilience in the face of class imbalance. Second, recognizing the quadratic complexity of dynamic routing leading to training inefficiencies, we adopt the attention mechanism as a more computationally efficient alternative. Our results demonstrate that CardioCaps surpasses traditional machine learning baseline methods, including Logistic Regression, Random Forest, and XGBoost with sampling methods and a class weight matrix. Furthermore, CardioCaps outperforms other deep learning baseline methods such as CNNs, ResNets, U-Nets, and ViTs, as well as advanced CapsNets methods such as EM-CapsNets and Efficient-CapsNets. Notably, our model demonstrates robustness to class imbalance, achieving high precision even in datasets with a substantial proportion of negative cases.
Large vision-language models (LVLMs), designed to interpret and respond to human instructions, occasionally generate hallucinated or harmful content due to inappropriate instructions. This study uses linear probing to shed light on the hidden knowledge at the output layers of LVLMs. We demonstrate that the logit distributions of the first tokens contain sufficient information to determine whether to respond to the instructions, including recognizing unanswerable visual questions, defending against jailbreaking attacks, and identifying deceptive questions. Such hidden knowledge is gradually lost in logits of subsequent tokens during response generation. Then, we illustrate a simple decoding strategy at the generation of the first token, effectively improving the generated content. In experiments, we find a few interesting insights: First, the CLIP model already contains a strong signal for solving these tasks, which indicates potential bias in the existing datasets. Second, we observe performance improvement by utilizing the first logit distributions on three additional tasks, including indicating uncertainty in math solving, mitigating hallucination, and image classification. Last, with the same training data, simply finetuning LVLMs improves models' performance but is still inferior to linear probing on these tasks.
Unlike traditional vision-only models, vision language models (VLMs) offer an intuitive way to access visual content through language prompting by combining a large language model (LLM) with a vision encoder. However, both the LLM and the vision encoder come with their own set of biases, cue preferences, and shortcuts, which have been rigorously studied in uni-modal models. A timely question is how such (potentially misaligned) biases and cue preferences behave under multi-modal fusion in VLMs. As a first step towards a better understanding, we investigate a particularly well-studied vision-only bias - the texture vs. shape bias and the dominance of local over global information. As expected, we find that VLMs inherit this bias to some extent from their vision encoders. Surprisingly, the multi-modality alone proves to have important effects on the model behavior, i.e., the joint training and the language querying change the way visual cues are processed. While this direct impact of language-informed training on a model's visual perception is intriguing, it raises further questions on our ability to actively steer a model's output so that its prediction is based on particular visual cues of the user's choice. Interestingly, VLMs have an inherent tendency to recognize objects based on shape information, which is different from what a plain vision encoder would do. Further active steering towards shape-based classifications through language prompts is however limited. In contrast, active VLM steering towards texture-based decisions through simple natural language prompts is often more successful. URL: https://github.com/paulgavrikov/vlm_shapebias
In spite of their huge success, transformer models remain difficult to scale in depth. In this work, we develop a unified signal propagation theory and provide formulae that govern the moments of the forward and backward signal through the transformer model. Our framework can be used to understand and mitigate vanishing/exploding gradients, rank collapse, and instability associated with high attention scores. We also propose DeepScaleLM, an initialization and scaling scheme that conserves unit output/gradient moments throughout the model, enabling the training of very deep models with 1000 layers. We find that transformer models could be much deeper - our deep models with fewer parameters outperform shallow models in Language Modeling, Speech Translation, and Image Classification, across encoder-only, decoder-only and encoder-decoder variants, for both Pre-LN and Post-LN transformers, for multiple datasets and model sizes. These improvements also translate into improved performance on downstream Question Answering tasks and improved robustness for Image Classification.
Deep learning and Convolutional Neural Networks (CNNs) have driven major transformations in diverse research areas. However, their limitations in handling low-frequency information present obstacles in certain tasks like interpreting global structures or managing smooth transition images. Despite the promising performance of transformer structures in numerous tasks, their intricate optimization complexities highlight the persistent need for refined CNN enhancements using limited resources. Responding to these complexities, we introduce a novel framework, the Multiscale Low-Frequency Memory (MLFM) Network, with the goal to harness the full potential of CNNs while keeping their complexity unchanged. The MLFM efficiently preserves low-frequency information, enhancing performance in targeted computer vision tasks. Central to our MLFM is the Low-Frequency Memory Unit (LFMU), which stores various low-frequency data and forms a parallel channel to the core network. A key advantage of MLFM is its seamless compatibility with various prevalent networks, requiring no alterations to their original core structure. Testing on ImageNet demonstrated substantial accuracy improvements in multiple 2D CNNs, including ResNet, MobileNet, EfficientNet, and ConvNeXt. Furthermore, we showcase MLFM's versatility beyond traditional image classification by successfully integrating it into image-to-image translation tasks, specifically in semantic segmentation networks like FCN and U-Net. In conclusion, our work signifies a pivotal stride in the journey of optimizing the efficacy and efficiency of CNNs with limited resources. This research builds upon the existing CNN foundations and paves the way for future advancements in computer vision. Our codes are available at https://github.com/AlphaWuSeu/ MLFM.
We present a parameter-efficient method for continual video question-answering (VidQA) learning. Our method, named DAM, uses the proposed Dynamic Adapter Merging to (i) mitigate catastrophic forgetting, (ii) enable efficient adaptation to continually arriving datasets, (iii) handle inputs from unknown datasets during inference, and (iv) enable knowledge sharing across similar dataset domains. Given a set of continually streaming VidQA datasets, we sequentially train dataset-specific adapters for each dataset while freezing the parameters of a large pretrained video-language backbone. During inference, given a video-question sample from an unknown domain, our method first uses the proposed non-parametric router function to compute a probability for each adapter, reflecting how relevant that adapter is to the current video-question input instance. Subsequently, the proposed dynamic adapter merging scheme aggregates all the adapter weights into a new adapter instance tailored for that particular test sample to compute the final VidQA prediction, mitigating the impact of inaccurate router predictions and facilitating knowledge sharing across domains. Our DAM model outperforms prior state-of-the-art continual learning approaches by 9.1% while exhibiting 1.9% less forgetting on 6 VidQA datasets spanning various domains. We further extend DAM to continual image classification and image QA and outperform prior methods by a large margin. The code is publicly available at: https://github.com/klauscc/DAM
Convolutional Neural Networks (CNNs) are nowadays the model of choice in Computer Vision, thanks to their ability to automatize the feature extraction process in visual tasks. However, the knowledge acquired during training is fully subsymbolic, and hence difficult to understand and explain to end users. In this paper, we propose a new technique called HOLMES (HOLonym-MEronym based Semantic inspection) that decomposes a label into a set of related concepts, and provides component-level explanations for an image classification model. Specifically, HOLMES leverages ontologies, web scraping and transfer learning to automatically construct meronym (parts)-based detectors for a given holonym (class). Then, it produces heatmaps at the meronym level and finally, by probing the holonym CNN with occluded images, it highlights the importance of each part on the classification output. Compared to state-of-the-art saliency methods, HOLMES takes a step further and provides information about both where and what the holonym CNN is looking at, without relying on densely annotated datasets and without forcing concepts to be associated to single computational units. Extensive experimental evaluation on different categories of objects (animals, tools and vehicles) shows the feasibility of our approach. On average, HOLMES explanations include at least two meronyms, and the ablation of a single meronym roughly halves the holonym model confidence. The resulting heatmaps were quantitatively evaluated using the deletion/insertion/preservation curves. All metrics were comparable to those achieved by GradCAM, while offering the advantage of further decomposing the heatmap in human-understandable concepts, thus highlighting both the relevance of meronyms to object classification, as well as HOLMES ability to capture it. The code is available at https://github.com/FrancesC0de/HOLMES.
Few-Shot Class-Incremental Learning (FSCIL) enables machine learning systems to expand their inference capabilities to new classes using only a few labeled examples, without forgetting the previously learned classes. Classical backpropagation-based learning and its variants are often unsuitable for battery-powered, memory-constrained systems at the extreme edge. In this work, we introduce Online Few-Shot Class-Incremental Learning (O-FSCIL), based on a lightweight model consisting of a pretrained and metalearned feature extractor and an expandable explicit memory storing the class prototypes. The architecture is pretrained with a novel feature orthogonality regularization and metalearned with a multi-margin loss. For learning a new class, our approach extends the explicit memory with novel class prototypes, while the remaining architecture is kept frozen. This allows learning previously unseen classes based on only a few examples with one single pass (hence online). O-FSCIL obtains an average accuracy of 68.62% on the FSCIL CIFAR100 benchmark, achieving state-of-the-art results. Tailored for ultra-low-power platforms, we implement O-FSCIL on the 60 mW GAP9 microcontroller, demonstrating online learning capabilities within just 12 mJ per new class.
Parameter-efficient transfer learning (PETL) is proposed as a cost-effective way to transfer pre-trained models to downstream tasks, avoiding the high cost of updating entire large-scale pre-trained models (LPMs). In this work, we present Fine-grained Prompt Tuning (FPT), a novel PETL method for medical image classification. FPT significantly reduces memory consumption compared to other PETL methods, especially in high-resolution input contexts. To achieve this, we first freeze the weights of the LPM and construct a learnable lightweight side network. The frozen LPM takes high-resolution images as input to extract fine-grained features, while the side network is fed low-resolution images to reduce memory usage. To allow the side network to access pre-trained knowledge, we introduce fine-grained prompts that summarize information from the LPM through a fusion module. Important tokens selection and preloading techniques are employed to further reduce training cost and memory requirements. We evaluate FPT on four medical datasets with varying sizes, modalities, and complexities. Experimental results demonstrate that FPT achieves comparable performance to fine-tuning the entire LPM while using only 1.8% of the learnable parameters and 13% of the memory costs of an encoder ViT-B model with a 512 x 512 input resolution.
Continual learning requires a model to adapt to ongoing changes in the data distribution, and often to the set of tasks to be performed. It is rare, however, that the data and task changes are completely unpredictable. Given a description of an overarching goal or data theme, which we call a realm, humans can often guess what concepts are associated with it. We show here that the combination of a large language model and an image generation model can similarly provide useful premonitions as to how a continual learning challenge might develop over time. We use the large language model to generate text descriptions of semantically related classes that might potentially appear in the data stream in future. These descriptions are then rendered using Stable Diffusion to generate new labelled image samples. The resulting synthetic dataset is employed for supervised pre-training, but is discarded prior to commencing continual learning, along with the pre-training classification head. We find that the backbone of our pre-trained networks can learn representations useful for the downstream continual learning problem, thus becoming a valuable input to any existing continual learning method. Although there are complexities arising from the domain gap between real and synthetic images, we show that pre-training models in this manner improves multiple Class Incremenal Learning (CIL) methods on fine-grained image classification benchmarks. Supporting code can be found at https://github.com/cl-premonition/premonition.
An important and unsolved problem in computer vision is to ensure that the algorithms are robust to changes in image domains. We address this problem in the scenario where we have access to images from the target domains but no annotations. Motivated by the challenges of the OOD-CV benchmark where we encounter real world Out-of-Domain (OOD) nuisances and occlusion, we introduce a novel Bayesian approach to OOD robustness for object classification. Our work extends Compositional Neural Networks (CompNets), which have been shown to be robust to occlusion but degrade badly when tested on OOD data. We exploit the fact that CompNets contain a generative head defined over feature vectors represented by von Mises-Fisher (vMF) kernels, which correspond roughly to object parts, and can be learned without supervision. We obverse that some vMF kernels are similar between different domains, while others are not. This enables us to learn a transitional dictionary of vMF kernels that are intermediate between the source and target domains and train the generative model on this dictionary using the annotations on the source domain, followed by iterative refinement. This approach, termed Unsupervised Generative Transition (UGT), performs very well in OOD scenarios even when occlusion is present. UGT is evaluated on different OOD benchmarks including the OOD-CV dataset, several popular datasets (e.g., ImageNet-C [9]), artificial image corruptions (including adding occluders), and synthetic-to-real domain transfer, and does well in all scenarios outperforming SOTA alternatives (e.g. up to 10% top-1 accuracy on Occluded OOD-CV dataset).
Data augmentation is a crucial regularization technique for deep neural networks, particularly in medical image classification. Mainstream data augmentation (DA) methods are usually applied at the image level. Due to the specificity and diversity of medical imaging, expertise is often required to design effective DA strategies, and improper augmentation operations can degrade model performance. Although automatic augmentation methods exist, they are computationally intensive. Semantic data augmentation can implemented by translating features in feature space. However, over-translation may violate the image label. To address these issues, we propose \emph{Bayesian Random Semantic Data Augmentation} (BSDA), a computationally efficient and handcraft-free feature-level DA method. BSDA uses variational Bayesian to estimate the distribution of the augmentable magnitudes, and then a sample from this distribution is added to the original features to perform semantic data augmentation. We performed experiments on nine 2D and five 3D medical image datasets. Experimental results show that BSDA outperforms current DA methods. Additionally, BSDA can be easily assembled into CNNs or Transformers as a plug-and-play module, improving the network's performance. The code is available online at \url{https://github.com/YaoyaoZhu19/BSDA}.
In the era of Foundation Models' (FMs) rising prominence in AI, our study addresses the challenge of biases in medical images while the model operates in black-box (e.g., using FM API), particularly spurious correlations between pixels and sensitive attributes. Traditional methods for bias mitigation face limitations due to the restricted access to web-hosted FMs and difficulties in addressing the underlying bias encoded within the FM API. We propose a D(ebiased) N(oise) E(diting) strategy, termed DNE, which generates DNE noise to mask such spurious correlation. DNE is capable of mitigating bias both within the FM API embedding and the images themselves. Furthermore, DNE is suitable for both white-box and black-box FM APIs, where we introduced G(reedy) (Z)eroth-O(rder) (GeZO) optimization for it when the gradient is inaccessible in black-box APIs. Our whole pipeline enables fairness-aware image editing that can be applied across various medical contexts without requiring direct model manipulation or significant computational resources. Our empirical results demonstrate the method's effectiveness in maintaining fairness and utility across different patient groups and diseases. In the era of AI-driven medicine, this work contributes to making healthcare diagnostics more equitable, showcasing a practical solution for bias mitigation in pre-trained image FMs. Our code is provided at https://github.com/ubc-tea/DNE-foundation-model-fairness.
Knowledge distillation is an attractive approach for learning compact deep neural networks, which learns a lightweight student model by distilling knowledge from a complex teacher model. Attention-based knowledge distillation is a specific form of intermediate feature-based knowledge distillation that uses attention mechanisms to encourage the student to better mimic the teacher. However, most of the previous attention-based distillation approaches perform attention in the spatial domain, which primarily affects local regions in the input image. This may not be sufficient when we need to capture the broader context or global information necessary for effective knowledge transfer. In frequency domain, since each frequency is determined from all pixels of the image in spatial domain, it can contain global information about the image. Inspired by the benefits of the frequency domain, we propose a novel module that functions as an attention mechanism in the frequency domain. The module consists of a learnable global filter that can adjust the frequencies of student's features under the guidance of the teacher's features, which encourages the student's features to have patterns similar to the teacher's features. We then propose an enhanced knowledge review-based distillation model by leveraging the proposed frequency attention module. The extensive experiments with various teacher and student architectures on image classification and object detection benchmark datasets show that the proposed approach outperforms other knowledge distillation methods.
Shortcut learning is a phenomenon where machine learning models prioritize learning simple, potentially misleading cues from data that do not generalize well beyond the training set. While existing research primarily investigates this in the realm of image classification, this study extends the exploration of shortcut learning into medical image segmentation. We demonstrate that clinical annotations such as calipers, and the combination of zero-padded convolutions and center-cropped training sets in the dataset can inadvertently serve as shortcuts, impacting segmentation accuracy. We identify and evaluate the shortcut learning on two different but common medical image segmentation tasks. In addition, we suggest strategies to mitigate the influence of shortcut learning and improve the generalizability of the segmentation models. By uncovering the presence and implications of shortcuts in medical image segmentation, we provide insights and methodologies for evaluating and overcoming this pervasive challenge and call for attention in the community for shortcuts in segmentation. Our code is public at https://github.com/nina-weng/shortcut_skinseg .
Recently, the growing capabilities of deep generative models have underscored their potential in enhancing image classification accuracy. However, existing methods often demand the generation of a disproportionately large number of images compared to the original dataset, while having only marginal improvements in accuracy. This computationally expensive and time-consuming process hampers the practicality of such approaches. In this paper, we propose to address the efficiency of image generation by focusing on the specific needs and characteristics of the model. With a central tenet of active learning, our method, named ActGen, takes a training-aware approach to image generation. It aims to create images akin to the challenging or misclassified samples encountered by the current model and incorporates these generated images into the training set to augment model performance. ActGen introduces an attentive image guidance technique, using real images as guides during the denoising process of a diffusion model. The model's attention on class prompt is leveraged to ensure the preservation of similar foreground object while diversifying the background. Furthermore, we introduce a gradient-based generation guidance method, which employs two losses to generate more challenging samples and prevent the generated images from being too similar to previously generated ones. Experimental results on the CIFAR and ImageNet datasets demonstrate that our method achieves better performance with a significantly reduced number of generated images. Code is available at https://github.com/hunto/ActGen.
Despite extensive diagnostics and debugging by developers, AI systems sometimes exhibit harmful unintended behaviors. Finding and fixing these is challenging because the attack surface is so large -- it is not tractable to exhaustively search for inputs that may elicit harmful behaviors. Red-teaming and adversarial training (AT) are commonly used to improve robustness, however, they empirically struggle to fix failure modes that differ from the attacks used during training. In this work, we utilize latent adversarial training (LAT) to defend against vulnerabilities without leveraging knowledge of what they are or using inputs that elicit them. LAT makes use of the compressed, abstract, and structured latent representations of concepts that the network actually uses for prediction. Here, we use it to defend against failure modes without examples that elicit them. Specifically, we use LAT to remove trojans and defend against held-out classes of adversarial attacks. We show in image classification, text classification, and text generation tasks that LAT usually improves both robustness to novel attacks and performance on clean data relative to AT. This suggests that LAT can be a promising tool for defending against failure modes that are not explicitly identified by developers.
In this paper, we study a defense against poisoned encoders in SSL called distillation, which is a defense used in supervised learning originally. Distillation aims to distill knowledge from a given model (a.k.a the teacher net) and transfer it to another (a.k.a the student net). Now, we use it to distill benign knowledge from poisoned pre-trained encoders and transfer it to a new encoder, resulting in a clean pre-trained encoder. In particular, we conduct an empirical study on the effectiveness and performance of distillation against poisoned encoders. Using two state-of-the-art backdoor attacks against pre-trained image encoders and four commonly used image classification datasets, our experimental results show that distillation can reduce attack success rate from 80.87% to 27.51% while suffering a 6.35% loss in accuracy. Moreover, we investigate the impact of three core components of distillation on performance: teacher net, student net, and distillation loss. By comparing 4 different teacher nets, 3 student nets, and 6 distillation losses, we find that fine-tuned teacher nets, warm-up-training-based student nets, and attention-based distillation loss perform best, respectively.