We carefully evaluate a number of algorithms for learning in a federated environment, and test their utility for a variety of image classification tasks. We consider many issues that have not been adequately considered before: whether learning over data sets that do not have diverse sets of images affects the results; whether to use a pre-trained feature extraction "backbone"; how to evaluate learner performance (we argue that classification accuracy is not enough), among others. Overall, across a wide variety of settings, we find that vertically decomposing a neural network seems to give the best results, and outperforms more standard reconciliation-used methods.
Large numbers of radiographic images are available in knee radiology practices which could be used for training of deep learning models for diagnosis of knee abnormalities. However, those images do not typically contain readily available labels due to limitations of human annotations. The purpose of our study was to develop an automated labeling approach that improves the image classification model to distinguish normal knee images from those with abnormalities or prior arthroplasty. The automated labeler was trained on a small set of labeled data to automatically label a much larger set of unlabeled data, further improving the image classification performance for knee radiographic diagnosis. We developed our approach using 7,382 patients and validated it on a separate set of 637 patients. The final image classification model, trained using both manually labeled and pseudo-labeled data, had the higher weighted average AUC (WAUC: 0.903) value and higher AUC-ROC values among all classes (normal AUC-ROC: 0.894; abnormal AUC-ROC: 0.896, arthroplasty AUC-ROC: 0.990) compared to the baseline model (WAUC=0.857; normal AUC-ROC: 0.842; abnormal AUC-ROC: 0.848, arthroplasty AUC-ROC: 0.987), trained using only manually labeled data. DeLong tests show that the improvement is significant on normal (p-value<0.002) and abnormal (p-value<0.001) images. Our findings demonstrated that the proposed automated labeling approach significantly improves the performance of image classification for radiographic knee diagnosis, allowing for facilitating patient care and curation of large knee datasets.
Deep neural networks (DNNs) have made great strides in pushing the state-of-the-art in several challenging domains. Recent studies reveal that they are prone to making overconfident predictions. This greatly reduces the overall trust in model predictions, especially in safety-critical applications. Early work in improving model calibration employs post-processing techniques which rely on limited parameters and require a hold-out set. Some recent train-time calibration methods, which involve all model parameters, can outperform the postprocessing methods. To this end, we propose a new train-time calibration method, which features a simple, plug-and-play auxiliary loss known as multi-class alignment of predictive mean confidence and predictive certainty (MACC). It is based on the observation that a model miscalibration is directly related to its predictive certainty, so a higher gap between the mean confidence and certainty amounts to a poor calibration both for in-distribution and out-of-distribution predictions. Armed with this insight, our proposed loss explicitly encourages a confident (or underconfident) model to also provide a low (or high) spread in the presoftmax distribution. Extensive experiments on ten challenging datasets, covering in-domain, out-domain, non-visual recognition and medical image classification scenarios, show that our method achieves state-of-the-art calibration performance for both in-domain and out-domain predictions. Our code and models will be publicly released.
Hyperspectral image classification is gaining popularity for high-precision vision tasks in remote sensing, thanks to their ability to capture visual information available in a wide continuum of spectra. Researchers have been working on automating Hyperspectral image classification, with recent efforts leveraging Vision-Transformers. However, most research models only spectra information and lacks attention to the locality (i.e., neighboring pixels), which may be not sufficiently discriminative, resulting in performance limitations. To address this, we present three contributions: i) We introduce the Hyperspectral Locality-aware Image TransformEr (HyLITE), a vision transformer that models both local and spectral information, ii) A novel regularization function that promotes the integration of local-to-global information, and iii) Our proposed approach outperforms competing baselines by a significant margin, achieving up to 10% gains in accuracy. The trained models and the code are available at HyLITE.
This paper presents our proposed approach that won the first prize at the ICLR competition on Hardware Aware Efficient Training. The challenge is to achieve the highest possible accuracy in an image classification task in less than 10 minutes. The training is done on a small dataset of 5000 images picked randomly from CIFAR-10 dataset. The evaluation is performed by the competition organizers on a secret dataset with 1000 images of the same size. Our approach includes applying a series of technique for improving the generalization of ResNet-9 including: sharpness aware optimization, label smoothing, gradient centralization, input patch whitening as well as metalearning based training. Our experiments show that the ResNet-9 can achieve the accuracy of 88% while trained only on a 10% subset of CIFAR-10 dataset in less than 10 minuets
In this paper, we propose a new self-supervised learning (SSL) method for representations that enable logic operations. Representation learning has been applied to various tasks, such as image generation and retrieval. The logical controllability of representations is important for these tasks. Although some methods have been shown to enable the intuitive control of representations using natural languages as the inputs, representation control via logic operations between representations has not been demonstrated. Some SSL methods using representation synthesis (e.g., elementwise mean and maximum operations) have been proposed, but the operations performed in these methods do not incorporate logic operations. In this work, we propose a logic-operable self-supervised representation learning method by replacing the existing representation synthesis with the OR operation on the probabilistic extension of many-valued logic. The representations comprise a set of feature-possession degrees, which are truth values indicating the presence or absence of each feature in the image, and realize the logic operations (e.g., OR and AND). Our method can generate a representation that has the features of both representations or only those features common to both representations. In addition, the expression of the ambiguous presence of a feature is realized by indicating the feature-possession degree by the probability distribution of truth values of the many-valued logic. We showed that our method performs competitively in single and multi-label classification tasks compared with prior SSL methods using synthetic representations. Moreover, experiments on image retrieval using MNIST and PascalVOC showed that the representations of our method can be operated by OR and AND operations.
The way we analyse clinical texts has undergone major changes over the last years. The introduction of language models such as BERT led to adaptations for the (bio)medical domain like PubMedBERT and ClinicalBERT. These models rely on large databases of archived medical documents. While performing well in terms of accuracy, both the lack of interpretability and limitations to transfer across languages limit their use in clinical setting. We introduce a novel light-weight graph-based embedding method specifically catering radiology reports. It takes into account the structure and composition of the report, while also connecting medical terms in the report through the multi-lingual SNOMED Clinical Terms knowledge base. The resulting graph embedding uncovers the underlying relationships among clinical terms, achieving a representation that is better understandable for clinicians and clinically more accurate, without reliance on large pre-training datasets. We show the use of this embedding on two tasks namely disease classification of X-ray reports and image classification. For disease classification our model is competitive with its BERT-based counterparts, while being magnitudes smaller in size and training data requirements. For image classification, we show the effectiveness of the graph embedding leveraging cross-modal knowledge transfer and show how this method is usable across different languages.
Internet of Things (IoT) sensors are nowadays heavily utilized in various real-world applications ranging from wearables to smart buildings passing by agrotechnology and health monitoring. With the huge amounts of data generated by these tiny devices, Deep Learning (DL) models have been extensively used to enhance them with intelligent processing. However, with the urge for smaller and more accurate devices, DL models became too heavy to deploy. It is thus necessary to incorporate the hardware's limited resources in the design process. Therefore, inspired by the human brain known for its efficiency and low power consumption, we propose a shallow bidirectional network based on predictive coding theory and dynamic early exiting for halting further computations when a performance threshold is surpassed. We achieve comparable accuracy to VGG-16 in image classification on CIFAR-10 with fewer parameters and less computational complexity.
Solving image classification tasks given small training datasets remains an open challenge for modern computer vision. Aggressive data augmentation and generative models are among the most straightforward approaches to overcoming the lack of data. However, the first fails to be agnostic to varying image domains, while the latter requires additional compute and careful design. In this work, we study alternative regularization strategies to push the limits of supervised learning on small image classification datasets. In particular, along with the model size and training schedule scaling, we employ a heuristic to select (semi) optimal learning rate and weight decay couples via the norm of model parameters. By training on only 1% of the original CIFAR-10 training set (i.e., 50 images per class) and testing on ciFAIR-10, a variant of the original CIFAR without duplicated images, we reach a test accuracy of 66.5%, on par with the best state-of-the-art methods.
Vision transformer (ViT) and its variants have swept through visual learning leaderboards and offer state-of-the-art accuracy in tasks such as image classification, object detection, and semantic segmentation by attending to different parts of the visual input and capturing long-range spatial dependencies. However, these models are large and computation-heavy. For instance, the recently proposed ViT-B model has 86M parameters making it impractical for deployment on resource-constrained devices. As a result, their deployment on mobile and edge scenarios is limited. In our work, we aim to take a step toward bringing vision transformers to the edge by utilizing popular model compression techniques such as distillation, pruning, and quantization. Our chosen application environment is an unmanned aerial vehicle (UAV) that is battery-powered and memory-constrained, carrying a single-board computer on the scale of an NVIDIA Jetson Nano with 4GB of RAM. On the other hand, the UAV requires high accuracy close to that of state-of-the-art ViTs to ensure safe object avoidance in autonomous navigation, or correct localization of humans in search-and-rescue. Inference latency should also be minimized given the application requirements. Hence, our target is to enable rapid inference of a vision transformer on an NVIDIA Jetson Nano (4GB) with minimal accuracy loss. This allows us to deploy ViTs on resource-constrained devices, opening up new possibilities in surveillance, environmental monitoring, etc. Our implementation is made available at https://github.com/chensy7/efficient-vit.
Adversarial attacks dramatically change the output of an otherwise accurate learning system using a seemingly inconsequential modification to a piece of input data. Paradoxically, empirical evidence indicates that even systems which are robust to large random perturbations of the input data remain susceptible to small, easily constructed, adversarial perturbations of their inputs. Here, we show that this may be seen as a fundamental feature of classifiers working with high dimensional input data. We introduce a simple generic and generalisable framework for which key behaviours observed in practical systems arise with high probability -- notably the simultaneous susceptibility of the (otherwise accurate) model to easily constructed adversarial attacks, and robustness to random perturbations of the input data. We confirm that the same phenomena are directly observed in practical neural networks trained on standard image classification problems, where even large additive random noise fails to trigger the adversarial instability of the network. A surprising takeaway is that even small margins separating a classifier's decision surface from training and testing data can hide adversarial susceptibility from being detected using randomly sampled perturbations. Counterintuitively, using additive noise during training or testing is therefore inefficient for eradicating or detecting adversarial examples, and more demanding adversarial training is required.
The paper proposes an efficient structure for enhancing the performance of mobile-friendly vision transformer with small computational overhead. The vision transformer (ViT) is very attractive in that it reaches outperforming results in image classification, compared to conventional convolutional neural networks (CNNs). Due to its need of high computational resources, MobileNet-based ViT models such as MobileViT-S have been developed. However, their performance cannot reach the original ViT model. The proposed structure relieves the above weakness by storing the information from early attention stages and reusing it in the final classifier. This paper is motivated by the idea that the data itself from early attention stages can have important meaning for the final classification. In order to reuse the early information in attention stages, the average pooling results of various scaled features from early attention stages are used to expand channels in the fully-connected layer of the final classifier. It is expected that the inductive bias introduced by the averaged features can enhance the final performance. Because the proposed structure only needs the average pooling of features from the attention stages and channel expansions in the final classifier, its computational and storage overheads are very small, keeping the benefits of low-cost MobileNet-based ViT (MobileViT). Compared with the original MobileViTs on the ImageNet dataset, the proposed ExMobileViT has noticeable accuracy enhancements, having only about 5% additional parameters.
Foundation models pre-trained on large-scale data have been widely witnessed to achieve success in various natural imaging downstream tasks. Parameter-efficient fine-tuning (PEFT) methods aim to adapt foundation models to new domains by updating only a small portion of parameters in order to reduce computational overhead. However, the effectiveness of these PEFT methods, especially in cross-domain few-shot scenarios, e.g., medical image analysis, has not been fully explored. In this work, we facilitate the study of the performance of PEFT when adapting foundation models to medical image classification tasks. Furthermore, to alleviate the limitations of prompt introducing ways and approximation capabilities on Transformer architectures of mainstream prompt tuning methods, we propose the Embedded Prompt Tuning (EPT) method by embedding prompt tokens into the expanded channels. We also find that there are anomalies in the feature space distribution of foundation models during pre-training process, and prompt tuning can help mitigate this negative impact. To explain this phenomenon, we also introduce a novel perspective to understand prompt tuning: Prompt tuning is a distribution calibrator. And we support it by analyzing patch-wise scaling and feature separation operations contained in EPT. Our experiments show that EPT outperforms several state-of-the-art fine-tuning methods by a significant margin on few-shot medical image classification tasks, and completes the fine-tuning process within highly competitive time, indicating EPT is an effective PEFT method. The source code is available at github.com/zuwenqiang/EPT.
Estimating the transferability of publicly available pretrained models to a target task has assumed an important place for transfer learning tasks in recent years. Existing efforts propose metrics that allow a user to choose one model from a pool of pre-trained models without having to fine-tune each model individually and identify one explicitly. With the growth in the number of available pre-trained models and the popularity of model ensembles, it also becomes essential to study the transferability of multiple-source models for a given target task. The few existing efforts study transferability in such multi-source ensemble settings using just the outputs of the classification layer and neglect possible domain or task mismatch. Moreover, they overlook the most important factor while selecting the source models, viz., the cohesiveness factor between them, which can impact the performance and confidence in the prediction of the ensemble. To address these gaps, we propose a novel Optimal tranSport-based suBmOdular tRaNsferability metric (OSBORN) to estimate the transferability of an ensemble of models to a downstream task. OSBORN collectively accounts for image domain difference, task difference, and cohesiveness of models in the ensemble to provide reliable estimates of transferability. We gauge the performance of OSBORN on both image classification and semantic segmentation tasks. Our setup includes 28 source datasets, 11 target datasets, 5 model architectures, and 2 pre-training methods. We benchmark our method against current state-of-the-art metrics MS-LEEP and E-LEEP, and outperform them consistently using the proposed approach.
Transformers have shown superior performance on various vision tasks. Their large receptive field endows Transformer models with higher representation power than their CNN counterparts. Nevertheless, simply enlarging the receptive field also raises several concerns. On the one hand, using dense attention in ViT leads to excessive memory and computational cost, and features can be influenced by irrelevant parts that are beyond the region of interests. On the other hand, the handcrafted attention adopted in PVT or Swin Transformer is data agnostic and may limit the ability to model long-range relations. To solve this dilemma, we propose a novel deformable multi-head attention module, where the positions of key and value pairs in self-attention are adaptively allocated in a data-dependent way. This flexible scheme enables the proposed deformable attention to dynamically focus on relevant regions while maintains the representation power of global attention. On this basis, we present Deformable Attention Transformer (DAT), a general vision backbone efficient and effective for visual recognition. We further build an enhanced version DAT++. Extensive experiments show that our DAT++ achieves state-of-the-art results on various visual recognition benchmarks, with 85.9% ImageNet accuracy, 54.5 and 47.0 MS-COCO instance segmentation mAP, and 51.5 ADE20K semantic segmentation mIoU.
Deep neural networks (DNNs) must cater to a variety of users with different performance needs and budgets, leading to the costly practice of training, storing, and maintaining numerous user/task-specific models. There are solutions in the literature to deal with single dynamic or many-in-one models instead of many individual networks; however, they suffer from significant drops in performance, lack of generalization across different model architectures or different dimensions (e.g. depth, width, attention blocks), heavy model search requirements during training, and training a limited number of sub-models. To address these limitations, we propose SortedNet, a generalized and scalable training solution to harness the inherent modularity of DNNs. Thanks to a generalized nested architecture (which we refer as \textit{sorted} architecture in this paper) with shared parameters and its novel update scheme combining random sub-model sampling and a new gradient accumulation mechanism, SortedNet enables the training of sub-models simultaneously along with the training of the main model (without any significant extra training or inference overhead), simplifies dynamic model selection, customizes deployment during inference, and reduces the model storage requirement significantly. The versatility and scalability of SortedNet are validated through various architectures and tasks, including LLaMA, BERT, RoBERTa (NLP tasks), ResNet and MobileNet (image classification) demonstrating its superiority over existing dynamic training methods. For example, we introduce a novel adaptive self-speculative approach based on sorted-training to accelerate large language models decoding. Moreover, SortedNet is able to train 160 sub-models at once, achieving at least 96\% of the original model's performance.
Fine-grained image recognition is a longstanding computer vision challenge that focuses on differentiating objects belonging to multiple subordinate categories within the same meta-category. Since images belonging to the same meta-category usually share similar visual appearances, mining discriminative visual cues is the key to distinguishing fine-grained categories. Although commonly used image-level data augmentation techniques have achieved great success in generic image classification problems, they are rarely applied in fine-grained scenarios, because their random editing-region behavior is prone to destroy the discriminative visual cues residing in the subtle regions. In this paper, we propose diversifying the training data at the feature-level to alleviate the discriminative region loss problem. Specifically, we produce diversified augmented samples by translating image features along semantically meaningful directions. The semantic directions are estimated with a covariance prediction network, which predicts a sample-wise covariance matrix to adapt to the large intra-class variation inherent in fine-grained images. Furthermore, the covariance prediction network is jointly optimized with the classification network in a meta-learning manner to alleviate the degenerate solution problem. Experiments on four competitive fine-grained recognition benchmarks (CUB-200-2011, Stanford Cars, FGVC Aircrafts, NABirds) demonstrate that our method significantly improves the generalization performance on several popular classification networks (e.g., ResNets, DenseNets, EfficientNets, RegNets and ViT). Combined with a recently proposed method, our semantic data augmentation approach achieves state-of-the-art performance on the CUB-200-2011 dataset. The source code will be released.
Computer vision models have known performance disparities across attributes such as gender and skin tone. This means during tasks such as classification and detection, model performance differs for certain classes based on the demographics of the people in the image. These disparities have been shown to exist, but until now there has not been a unified approach to measure these differences for common use-cases of computer vision models. We present a new benchmark named FACET (FAirness in Computer Vision EvaluaTion), a large, publicly available evaluation set of 32k images for some of the most common vision tasks - image classification, object detection and segmentation. For every image in FACET, we hired expert reviewers to manually annotate person-related attributes such as perceived skin tone and hair type, manually draw bounding boxes and label fine-grained person-related classes such as disk jockey or guitarist. In addition, we use FACET to benchmark state-of-the-art vision models and present a deeper understanding of potential performance disparities and challenges across sensitive demographic attributes. With the exhaustive annotations collected, we probe models using single demographics attributes as well as multiple attributes using an intersectional approach (e.g. hair color and perceived skin tone). Our results show that classification, detection, segmentation, and visual grounding models exhibit performance disparities across demographic attributes and intersections of attributes. These harms suggest that not all people represented in datasets receive fair and equitable treatment in these vision tasks. We hope current and future results using our benchmark will contribute to fairer, more robust vision models. FACET is available publicly at https://facet.metademolab.com/
In energy-efficient schemes, finding the optimal size of deep learning models is very important and has a broad impact. Meanwhile, recent studies have reported an unexpected phenomenon, the sparse double descent: as the model's sparsity increases, the performance first worsens, then improves, and finally deteriorates. Such a non-monotonic behavior raises serious questions about the optimal model's size to maintain high performance: the model needs to be sufficiently over-parametrized, but having too many parameters wastes training resources. In this paper, we aim to find the best trade-off efficiently. More precisely, we tackle the occurrence of the sparse double descent and present some solutions to avoid it. Firstly, we show that a simple $\ell_2$ regularization method can help to mitigate this phenomenon but sacrifices the performance/sparsity compromise. To overcome this problem, we then introduce a learning scheme in which distilling knowledge regularizes the student model. Supported by experimental results achieved using typical image classification setups, we show that this approach leads to the avoidance of such a phenomenon.
Deep learning (DL)-based solutions have been extensively researched in the medical domain in recent years, enhancing the efficacy of diagnosis, planning, and treatment. Since the usage of health-related data is strictly regulated, processing medical records outside the hospital environment for developing and using DL models demands robust data protection measures. At the same time, it can be challenging to guarantee that a DL solution delivers a minimum level of performance when being trained on secured data, without being specifically designed for the given task. Our approach uses singular value decomposition (SVD) and principal component analysis (PCA) to obfuscate the medical images before employing them in the DL analysis. The capability of DL algorithms to extract relevant information from secured data is assessed on a task of angiographic view classification based on obfuscated frames. The security level is probed by simulated artificial intelligence (AI)-based reconstruction attacks, considering two threat actors with different prior knowledge of the targeted data. The degree of privacy is quantitatively measured using similarity indices. Although a trade-off between privacy and accuracy should be considered, the proposed technique allows for training the angiographic view classifier exclusively on secured data with satisfactory performance and with no computational overhead, model adaptation, or hyperparameter tuning. While the obfuscated medical image content is well protected against human perception, the hypothetical reconstruction attack proved that it is also difficult to recover the complete information of the original frames.