It is a challenging task to learn discriminative representation from images and videos, due to large local redundancy and complex global dependency in these visual data. Convolution neural networks (CNNs) and vision transformers (ViTs) have been two dominant frameworks in the past few years. Though CNNs can efficiently decrease local redundancy by convolution within a small neighborhood, the limited receptive field makes it hard to capture global dependency. Alternatively, ViTs can effectively capture long-range dependency via self-attention, while blind similarity comparisons among all the tokens lead to high redundancy. To resolve these problems, we propose a novel Unified transFormer (UniFormer), which can seamlessly integrate the merits of convolution and self-attention in a concise transformer format. Different from the typical transformer blocks, the relation aggregators in our UniFormer block are equipped with local and global token affinity respectively in shallow and deep layers, allowing to tackle both redundancy and dependency for efficient and effective representation learning. Finally, we flexibly stack our UniFormer blocks into a new powerful backbone, and adopt it for various vision tasks from image to video domain, from classification to dense prediction. Without any extra training data, our UniFormer achieves 86.3 top-1 accuracy on ImageNet-1K classification. With only ImageNet-1K pre-training, it can simply achieve state-of-the-art performance in a broad range of downstream tasks, e.g., it obtains 82.9/84.8 top-1 accuracy on Kinetics-400/600, 60.9/71.2 top-1 accuracy on Sth-Sth V1/V2 video classification, 53.8 box AP and 46.4 mask AP on COCO object detection, 50.8 mIoU on ADE20K semantic segmentation, and 77.4 AP on COCO pose estimation. We further build an efficient UniFormer with 2-4x higher throughput. Code is available at https://github.com/Sense-X/UniFormer.
Recent works on Multimodal 3D Computer-aided diagnosis have demonstrated that obtaining a competitive automatic diagnosis model when a 3D convolution neural network (CNN) brings more parameters and medical images are scarce remains nontrivial and challenging. Considering both consistencies of regions of interest in multimodal images and diagnostic accuracy, we propose a novel mutual attention-based hybrid dimensional network for MultiModal 3D medical image classification (MMNet). The hybrid dimensional network integrates 2D CNN with 3D convolution modules to generate deeper and more informative feature maps, and reduce the training complexity of 3D fusion. Besides, the pre-trained model of ImageNet can be used in 2D CNN, which improves the performance of the model. The stereoscopic attention is focused on building rich contextual interdependencies of the region in 3D medical images. To improve the regional correlation of pathological tissues in multimodal medical images, we further design a mutual attention framework in the network to build the region-wise consistency in similar stereoscopic regions of different image modalities, providing an implicit manner to instruct the network to focus on pathological tissues. MMNet outperforms many previous solutions and achieves results competitive to the state-of-the-art on three multimodal imaging datasets, i.e., Parotid Gland Tumor (PGT) dataset, the MRNet dataset, and the PROSTATEx dataset, and its advantages are validated by extensive experiments.
The feature learning methods based on convolutional neural network (CNN) have successfully produced tremendous achievements in image classification tasks. However, the inherent noise and some other factors may weaken the effectiveness of the convolutional feature statistics. In this paper, we investigate Discrete Wavelet Transform (DWT) in the frequency domain and design a new Wavelet-Attention (WA) block to only implement attention in the high-frequency domain. Based on this, we propose a Wavelet-Attention convolutional neural network (WA-CNN) for image classification. Specifically, WA-CNN decomposes the feature maps into low-frequency and high-frequency components for storing the structures of the basic objects, as well as the detailed information and noise, respectively. Then, the WA block is leveraged to capture the detailed information in the high-frequency domain with different attention factors but reserves the basic object structures in the low-frequency domain. Experimental results on CIFAR-10 and CIFAR-100 datasets show that our proposed WA-CNN achieves significant improvements in classification accuracy compared to other related networks. Specifically, based on MobileNetV2 backbones, WA-CNN achieves 1.26% Top-1 accuracy improvement on the CIFAR-10 benchmark and 1.54% Top-1 accuracy improvement on the CIFAR-100 benchmark.
The learning process of deep learning methods usually updates the model's parameters in multiple iterations. Each iteration can be viewed as the first-order approximation of Taylor's series expansion. The remainder, which consists of higher-order terms, is usually ignored in the learning process for simplicity. This learning scheme empowers various multimedia based applications, such as image retrieval, recommendation system, and video search. Generally, multimedia data (e.g., images) are semantics-rich and high-dimensional, hence the remainders of approximations are possibly non-zero. In this work, we consider the remainder to be informative and study how it affects the learning process. To this end, we propose a new learning approach, namely gradient adjustment learning (GAL), to leverage the knowledge learned from the past training iterations to adjust vanilla gradients, such that the remainders are minimized and the approximations are improved. The proposed GAL is model- and optimizer-agnostic, and is easy to adapt to the standard learning framework. It is evaluated on three tasks, i.e., image classification, object detection, and regression, with state-of-the-art models and optimizers. The experiments show that the proposed GAL consistently enhances the evaluated models, whereas the ablation studies validate various aspects of the proposed GAL. The code is available at \url{https://github.com/luoyan407/gradient_adjustment.git}.
Generative Adversarial Networks (GANs) have shown promise in augmenting datasets and boosting convolutional neural networks' (CNN) performance on image classification tasks. But they introduce more hyperparameters to tune as well as the need for additional time and computational power to train supplementary to the CNN. In this work, we examine the potential for Auxiliary-Classifier GANs (AC-GANs) as a 'one-stop-shop' architecture for image classification, particularly in low data regimes. Additionally, we explore modifications to the typical AC-GAN framework, changing the generator's latent space sampling scheme and employing a Wasserstein loss with gradient penalty to stabilize the simultaneous training of image synthesis and classification. Through experiments on images of varying resolutions and complexity, we demonstrate that AC-GANs show promise in image classification, achieving competitive performance with standard CNNs. These methods can be employed as an 'all-in-one' framework with particular utility in the absence of large amounts of training data.
Although convolutional networks have been the dominant architecture for vision tasks for many years, recent experiments have shown that Transformer-based models, most notably the Vision Transformer (ViT), may exceed their performance in some settings. However, due to the quadratic runtime of the self-attention layers in Transformers, ViTs require the use of patch embeddings, which group together small regions of the image into single input features, in order to be applied to larger image sizes. This raises a question: Is the performance of ViTs due to the inherently-more-powerful Transformer architecture, or is it at least partly due to using patches as the input representation? In this paper, we present some evidence for the latter: specifically, we propose the ConvMixer, an extremely simple model that is similar in spirit to the ViT and the even-more-basic MLP-Mixer in that it operates directly on patches as input, separates the mixing of spatial and channel dimensions, and maintains equal size and resolution throughout the network. In contrast, however, the ConvMixer uses only standard convolutions to achieve the mixing steps. Despite its simplicity, we show that the ConvMixer outperforms the ViT, MLP-Mixer, and some of their variants for similar parameter counts and data set sizes, in addition to outperforming classical vision models such as the ResNet. Our code is available at https://github.com/locuslab/convmixer.
A problem with Convolutional Neural Networks (CNNs) is that they require large datasets to obtain adequate robustness; on small datasets, they are prone to overfitting. Many methods have been proposed to overcome this shortcoming with CNNs. In cases where additional samples cannot easily be collected, a common approach is to generate more data points from existing data using an augmentation technique. In image classification, many augmentation approaches utilize simple image manipulation algorithms. In this work, we build ensembles on the data level by adding images generated by combining fourteen augmentation approaches, three of which are proposed here for the first time. These novel methods are based on the Fourier Transform (FT), the Radon Transform (RT) and the Discrete Cosine Transform (DCT). Pretrained ResNet50 networks are finetuned on training sets that include images derived from each augmentation method. These networks and several fusions are evaluated and compared across eleven benchmarks. Results show that building ensembles on the data level by combining different data augmentation methods produce classifiers that not only compete competitively against the state-of-the-art but often surpass the best approaches reported in the literature.
The article presents a new multi-label comprehensive image dataset from flexible endoscopy, colonoscopy and capsule endoscopy, named ERS. The collection has been labeled according to the full medical specification of 'Minimum Standard Terminology 3.0' (MST 3.0), describing all possible findings in the gastrointestinal tract (104 possible labels), extended with an additional 19 labels useful in common machine learning applications. The dataset contains around 6000 precisely and 115,000 approximately labeled frames from endoscopy videos, 3600 precise and 22,600 approximate segmentation masks, and 1.23 million unlabeled frames from flexible and capsule endoscopy videos. The labeled data cover almost entirely the MST 3.0 standard. The data came from 1520 videos of 1135 patients. Additionally, this paper proposes and describes four exemplary experiments in gastrointestinal image classification task performed using the created dataset. The obtained results indicate the high usefulness and flexibility of the dataset in training and testing machine learning algorithms in the field of endoscopic data analysis.
Deep learning models have been shown to be vulnerable to recent backdoor attacks. A backdoored model behaves normally for inputs containing no attacker-secretly-chosen trigger and maliciously for inputs with the trigger. To date, backdoor attacks and countermeasures mainly focus on image classification tasks. And most of them are implemented in the digital world with digital triggers. Besides the classification tasks, object detection systems are also considered as one of the basic foundations of computer vision tasks. However, there is no investigation and understanding of the backdoor vulnerability of the object detector, even in the digital world with digital triggers. For the first time, this work demonstrates that existing object detectors are inherently susceptible to physical backdoor attacks. We use a natural T-shirt bought from a market as a trigger to enable the cloaking effect--the person bounding-box disappears in front of the object detector. We show that such a backdoor can be implanted from two exploitable attack scenarios into the object detector, which is outsourced or fine-tuned through a pretrained model. We have extensively evaluated three popular object detection algorithms: anchor-based Yolo-V3, Yolo-V4, and anchor-free CenterNet. Building upon 19 videos shot in real-world scenes, we confirm that the backdoor attack is robust against various factors: movement, distance, angle, non-rigid deformation, and lighting. Specifically, the attack success rate (ASR) in most videos is 100% or close to it, while the clean data accuracy of the backdoored model is the same as its clean counterpart. The latter implies that it is infeasible to detect the backdoor behavior merely through a validation set. The averaged ASR still remains sufficiently high to be 78% in the transfer learning attack scenarios evaluated on CenterNet. See the demo video on https://youtu.be/Q3HOF4OobbY.
Prior work has studied different visual modalities in isolation and developed separate architectures for recognition of images, videos, and 3D data. Instead, in this paper, we propose a single model which excels at classifying images, videos, and single-view 3D data using exactly the same model parameters. Our 'Omnivore' model leverages the flexibility of transformer-based architectures and is trained jointly on classification tasks from different modalities. Omnivore is simple to train, uses off-the-shelf standard datasets, and performs at-par or better than modality-specific models of the same size. A single Omnivore model obtains 86.0% on ImageNet, 84.1% on Kinetics, and 67.1% on SUN RGB-D. After finetuning, our models outperform prior work on a variety of vision tasks and generalize across modalities. Omnivore's shared visual representation naturally enables cross-modal recognition without access to correspondences between modalities. We hope our results motivate researchers to model visual modalities together.
Model pre-training is a cornerstone of modern visual recognition systems. Although fully supervised pre-training on datasets like ImageNet is still the de-facto standard, recent studies suggest that large-scale weakly supervised pre-training can outperform fully supervised approaches. This paper revisits weakly-supervised pre-training of models using hashtag supervision with modern versions of residual networks and the largest-ever dataset of images and corresponding hashtags. We study the performance of the resulting models in various transfer-learning settings including zero-shot transfer. We also compare our models with those obtained via large-scale self-supervised learning. We find our weakly-supervised models to be very competitive across all settings, and find they substantially outperform their self-supervised counterparts. We also include an investigation into whether our models learned potentially troubling associations or stereotypes. Overall, our results provide a compelling argument for the use of weakly supervised learning in the development of visual recognition systems. Our models, Supervised Weakly through hashtAGs (SWAG), are available publicly.
Labeling a large set of data is expensive. Active learning aims to tackle this problem by asking to annotate only the most informative data from the unlabeled set. We propose a novel active learning approach that utilizes self-supervised pretext tasks and a unique data sampler to select data that are both difficult and representative. We discover that the loss of a simple self-supervised pretext task, such as rotation prediction, is closely correlated to the downstream task loss. Before the active learning iterations, the pretext task learner is trained on the unlabeled set, and the unlabeled data are sorted and split into batches by their pretext task losses. In each active learning iteration, the main task model is used to sample the most uncertain data in a batch to be annotated. We evaluate our method on various image classification and segmentation benchmarks and achieve compelling performances on CIFAR10, Caltech-101, ImageNet, and Cityscapes. We further show that our method performs well on imbalanced datasets, and can be an effective solution to the cold-start problem where active learning performance is affected by the randomly sampled initial labeled set.
Image classification is widely used to build predictive models for breast cancer diagnosis. Most existing approaches overwhelmingly rely on deep convolutional networks to build such diagnosis pipelines. These model architectures, although remarkable in performance, are black-box systems that provide minimal insight into the inner logic behind their predictions. This is a major drawback as the explainability of prediction is vital for applications such as cancer diagnosis. In this paper, we address this issue by proposing an explainable machine learning pipeline for breast cancer diagnosis based on ultrasound images. We extract first- and second-order texture features of the ultrasound images and use them to build a probabilistic ensemble of decision tree classifiers. Each decision tree learns to classify the input ultrasound image by learning a set of robust decision thresholds for texture features of the image. The decision path of the model predictions can then be interpreted by decomposing the learned decision trees. Our results show that our proposed framework achieves high predictive performance while being explainable.
Compressing deep networks is essential to expand their range of applications to constrained settings. The need for compression however often arises long after the model was trained, when the original data might no longer be available. On the other hand, unlabeled data, not necessarily related to the target task, is usually plentiful, especially in image classification tasks. In this work, we propose a scheme to leverage such samples to distill the knowledge learned by a large teacher network to a smaller student. The proposed technique relies on (i) preferentially sampling datapoints that appear related, and (ii) taking better advantage of the learning signal. We show that the former speeds up the student's convergence, while the latter boosts its performance, achieving performances closed to what can be expected with the original data.
Due to diverse architectures in deep neural networks (DNNs) with severe overparameterization, regularization techniques are critical for finding optimal solutions in the huge hypothesis space. In this paper, we propose an effective regularization technique, called Neighborhood Region Smoothing (NRS). NRS leverages the finding that models would benefit from converging to flat minima, and tries to regularize the neighborhood region in weight space to yield approximate outputs. Specifically, gap between outputs of models in the neighborhood region is gauged by a defined metric based on Kullback-Leibler divergence. This metric provides similar insights with the minimum description length principle on interpreting flat minima. By minimizing both this divergence and empirical loss, NRS could explicitly drive the optimizer towards converging to flat minima. We confirm the effectiveness of NRS by performing image classification tasks across a wide range of model architectures on commonly-used datasets such as CIFAR and ImageNet, where generalization ability could be universally improved. Also, we empirically show that the minima found by NRS would have relatively smaller Hessian eigenvalues compared to the conventional method, which is considered as the evidence of flat minima.
Machine learning (ML) methods are used in most technical areas such as image recognition, product recommendation, financial analysis, medical diagnosis, and predictive maintenance. An important aspect of implementing ML methods involves controlling the learning process for the ML method so as to maximize the performance of the method under consideration. Hyperparameter tuning is the process of selecting a suitable set of ML method parameters that control its learning process. In this work, we demonstrate the use of discrete simulation optimization methods such as ranking and selection (R&S) and random search for identifying a hyperparameter set that maximizes the performance of a ML method. Specifically, we use the KN R&S method and the stochastic ruler random search method and one of its variations for this purpose. We also construct the theoretical basis for applying the KN method, which determines the optimal solution with a statistical guarantee via solution space enumeration. In comparison, the stochastic ruler method asymptotically converges to global optima and incurs smaller computational overheads. We demonstrate the application of these methods to a wide variety of machine learning models, including deep neural network models used for time series prediction and image classification. We benchmark our application of these methods with state-of-the-art hyperparameter optimization libraries such as $hyperopt$ and $mango$. The KN method consistently outperforms $hyperopt$'s random search (RS) and Tree of Parzen Estimators (TPE) methods. The stochastic ruler method outperforms the $hyperopt$ RS method and offers statistically comparable performance with respect to $hyperopt$'s TPE method and the $mango$ algorithm.
Neural ordinary differential equations (NODEs) -- parametrizations of differential equations using neural networks -- have shown tremendous promise in learning models of unknown continuous-time dynamical systems from data. However, every forward evaluation of a NODE requires numerical integration of the neural network used to capture the system dynamics, making their training prohibitively expensive. Existing works rely on off-the-shelf adaptive step-size numerical integration schemes, which often require an excessive number of evaluations of the underlying dynamics network to obtain sufficient accuracy for training. By contrast, we accelerate the evaluation and the training of NODEs by proposing a data-driven approach to their numerical integration. The proposed Taylor-Lagrange NODEs (TL-NODEs) use a fixed-order Taylor expansion for numerical integration, while also learning to estimate the expansion's approximation error. As a result, the proposed approach achieves the same accuracy as adaptive step-size schemes while employing only low-order Taylor expansions, thus greatly reducing the computational cost necessary to integrate the NODE. A suite of numerical experiments, including modeling dynamical systems, image classification, and density estimation, demonstrate that TL-NODEs can be trained more than an order of magnitude faster than state-of-the-art approaches, without any loss in performance.
With this work we are explaining the "You Only Look Once" (YOLO) single-stage object detection approach as a parallel classification of 10647 fixed region proposals. We support this view by showing that each of YOLOs output pixel is attentive to a specific sub-region of previous layers, comparable to a local region proposal. This understanding reduces the conceptual gap between YOLO-like single-stage object detection models, RCNN-like two-stage region proposal based models, and ResNet-like image classification models. In addition, we created interactive exploration tools for a better visual understanding of the YOLO information processing streams: https://limchr.github.io/yolo_visualization
We propose a Multi-level Second-order (MlSo) few-shot learning network for supervised or unsupervised few-shot image classification and few-shot action recognition. We leverage so-called power-normalized second-order base learner streams combined with features that express multiple levels of visual abstraction, and we use self-supervised discriminating mechanisms. As Second-order Pooling (SoP) is popular in image recognition, we employ its basic element-wise variant in our pipeline. The goal of multi-level feature design is to extract feature representations at different layer-wise levels of CNN, realizing several levels of visual abstraction to achieve robust few-shot learning. As SoP can handle convolutional feature maps of varying spatial sizes, we also introduce image inputs at multiple spatial scales into MlSo. To exploit the discriminative information from multi-level and multi-scale features, we develop a Feature Matching (FM) module that reweights their respective branches. We also introduce a self-supervised step, which is a discriminator of the spatial level and the scale of abstraction. Our pipeline is trained in an end-to-end manner. With a simple architecture, we demonstrate respectable results on standard datasets such as Omniglot, mini-ImageNet, tiered-ImageNet, Open MIC, fine-grained datasets such as CUB Birds, Stanford Dogs and Cars, and action recognition datasets such as HMDB51, UCF101, and mini-MIT.
Representation knowledge distillation aims at transferring rich information from one model to another. Common approaches for representation distillation mainly focus on the direct minimization of distance metrics between the models' embedding vectors. Such direct methods may be limited in transferring high-order dependencies embedded in the representation vectors, or in handling the capacity gap between the teacher and student models. Moreover, in standard knowledge distillation, the teacher is trained without awareness of the student's characteristics and capacity. In this paper, we explore two mechanisms for enhancing representation distillation using classifier sharing between the teacher and student. We first investigate a simple scheme where the teacher's classifier is connected to the student backbone, acting as an additional classification head. Then, we propose a student-aware mechanism that asks to tailor the teacher model to a student with limited capacity by training the teacher with a temporary student's head. We analyze and compare these two mechanisms and show their effectiveness on various datasets and tasks, including image classification, fine-grained classification, and face verification. In particular, we achieve state-of-the-art results for face verification on the IJB-C dataset for a MobileFaceNet model: TAR@(FAR=1e-5)=93.7\%. Code is available at https://github.com/Alibaba-MIIL/HeadSharingKD.