When designing Convolutional Neural Networks (CNNs), one must select the size\break of the convolutional kernels before training. Recent works show CNNs benefit from different kernel sizes at different layers, but exploring all possible combinations is unfeasible in practice. A more efficient approach is to learn the kernel size during training. However, existing works that learn the kernel size have a limited bandwidth. These approaches scale kernels by dilation, and thus the detail they can describe is limited. In this work, we propose FlexConv, a novel convolutional operation with which high bandwidth convolutional kernels of learnable kernel size can be learned at a fixed parameter cost. FlexNets model long-term dependencies without the use of pooling, achieve state-of-the-art performance on several sequential datasets, outperform recent works with learned kernel sizes, and are competitive with much deeper ResNets on image benchmark datasets. Additionally, FlexNets can be deployed at higher resolutions than those seen during training. To avoid aliasing, we propose a novel kernel parameterization with which the frequency of the kernels can be analytically controlled. Our novel kernel parameterization shows higher descriptive power and faster convergence speed than existing parameterizations. This leads to important improvements in classification accuracy.
Due to the vulnerability of deep neural networks (DNNs) to adversarial examples, a large number of defense techniques have been proposed to alleviate this problem in recent years. However, the progress of building more robust models is usually hampered by the incomplete or incorrect robustness evaluation. To accelerate the research on reliable evaluation of adversarial robustness of the current defense models in image classification, the TSAIL group at Tsinghua University and the Alibaba Security group organized this competition along with a CVPR 2021 workshop on adversarial machine learning (https://aisecure-workshop.github.io/amlcvpr2021/). The purpose of this competition is to motivate novel attack algorithms to evaluate adversarial robustness more effectively and reliably. The participants were encouraged to develop stronger white-box attack algorithms to find the worst-case robustness of different defenses. This competition was conducted on an adversarial robustness evaluation platform -- ARES (https://github.com/thu-ml/ares), and is held on the TianChi platform (https://tianchi.aliyun.com/competition/entrance/531847/introduction) as one of the series of AI Security Challengers Program. After the competition, we summarized the results and established a new adversarial robustness benchmark at https://ml.cs.tsinghua.edu.cn/ares-bench/, which allows users to upload adversarial attack algorithms and defense models for evaluation.
Depth is the hallmark of deep neural networks. But more depth means more sequential computation and higher latency. This begs the question -- is it possible to build high-performing "non-deep" neural networks? We show that it is. To do so, we use parallel subnetworks instead of stacking one layer after another. This helps effectively reduce depth while maintaining high performance. By utilizing parallel substructures, we show, for the first time, that a network with a depth of just 12 can achieve top-1 accuracy over 80% on ImageNet, 96% on CIFAR10, and 81% on CIFAR100. We also show that a network with a low-depth (12) backbone can achieve an AP of 48% on MS-COCO. We analyze the scaling rules for our design and show how to increase performance without changing the network's depth. Finally, we provide a proof of concept for how non-deep networks could be used to build low-latency recognition systems. Code is available at https://github.com/imankgoyal/NonDeepNetworks.
We present TWIST, a simple and theoretically explainable self-supervised representation learning method by classifying large-scale unlabeled datasets in an end-to-end way. We employ a siamese network terminated by a softmax operation to produce twin class distributions of two augmented images. Without supervision, we enforce the class distributions of different augmentations to be consistent. However, simply minimizing the divergence between augmentations will cause collapsed solutions, i.e., outputting the same class probability distribution for all images. In this case, no information about the input image is left. To solve this problem, we propose to maximize the mutual information between the input and the class predictions. Specifically, we minimize the entropy of the distribution for each sample to make the class prediction for each sample assertive and maximize the entropy of the mean distribution to make the predictions of different samples diverse. In this way, TWIST can naturally avoid the collapsed solutions without specific designs such as asymmetric network, stop-gradient operation, or momentum encoder. As a result, TWIST outperforms state-of-the-art methods on a wide range of tasks. Especially, TWIST performs surprisingly well on semi-supervised learning, achieving 61.2% top-1 accuracy with 1% ImageNet labels using a ResNet-50 as backbone, surpassing previous best results by an absolute improvement of 6.2%. Codes and pre-trained models are given on: https://github.com/bytedance/TWIST
Nowadays, most classification networks use one-hot encoding to represent categorical data because of its simplicity. However, one-hot encoding may affect the generalization ability as it neglects inter-class correlations. We observe that, even when a neural network trained with one-hot labels produces incorrect predictions, it still pays attention to the target image region and reveals which classes confuse the network. Inspired by this observation, we propose a confusion-focusing mechanism to address the class-confusion issue. Our confusion-focusing mechanism is implemented by a two-branch network architecture. Its baseline branch generates confusing classes, and its FocusNet branch, whose architecture is flexible, discriminates correct labels from these confusing classes. We also introduce a novel focus-picking loss function to improve classification accuracy by encouraging FocusNet to focus on the most confusing classes. The experimental results validate that our FocusNet is effective for image classification on common datasets, and that our focus-picking loss function can also benefit the current neural networks in improving their classification accuracy.
Few-shot class incremental learning -- the problem of updating a trained classifier to discriminate among an expanded set of classes with limited labeled data -- is a key challenge for machine learning systems deployed in non-stationary environments. Existing approaches to the problem rely on complex model architectures and training procedures that are difficult to tune and re-use. In this paper, we present an extremely simple approach that enables the use of ordinary logistic regression classifiers for few-shot incremental learning. The key to this approach is a new family of subspace regularization schemes that encourage weight vectors for new classes to lie close to the subspace spanned by the weights of existing classes. When combined with pretrained convolutional feature extractors, logistic regression models trained with subspace regularization outperform specialized, state-of-the-art approaches to few-shot incremental image classification by up to 22% on the miniImageNet dataset. Because of its simplicity, subspace regularization can be straightforwardly extended to incorporate additional background information about the new classes (including class names and descriptions specified in natural language); these further improve accuracy by up to 2%. Our results show that simple geometric regularization of class representations offers an effective tool for continual learning.
Empirical science of neural scaling laws is a rapidly growing area of significant importance to the future of machine learning, particularly in the light of recent breakthroughs achieved by large-scale pre-trained models such as GPT-3, CLIP and DALL-e. Accurately predicting the neural network performance with increasing resources such as data, compute and model size provides a more comprehensive evaluation of different approaches across multiple scales, as opposed to traditional point-wise comparisons of fixed-size models on fixed-size benchmarks, and, most importantly, allows for focus on the best-scaling, and thus most promising in the future, approaches. In this work, we consider a challenging problem of few-shot learning in image classification, especially when the target data distribution in the few-shot phase is different from the source, training, data distribution, in a sense that it includes new image classes not encountered during training. Our current main goal is to investigate how the amount of pre-training data affects the few-shot generalization performance of standard image classifiers. Our key observations are that (1) such performance improvements are well-approximated by power laws (linear log-log plots) as the training set size increases, (2) this applies to both cases of target data coming from either the same or from a different domain (i.e., new classes) as the training data, and (3) few-shot performance on new classes converges at a faster rate than the standard classification performance on previously seen classes. Our findings shed new light on the relationship between scale and generalization.
Classification of images within the compressed domain offers significant benefits. These benefits include reduced memory and computational requirements of a classification system. This paper proposes two such methods as a proof of concept: The first classifies within the JPEG image transform domain (i.e. DCT transform data); the second classifies the JPEG compressed binary bitstream directly. These two methods are implemented using Residual Network CNNs and an adapted Vision Transformer. Top-1 accuracy of approximately 70% and 60% were achieved using these methods respectively when classifying the Caltech C101 database. Although these results are significantly behind the state of the art for classification for this database (~95%), it illustrates the first time direct bitstream image classification has been achieved. This work confirms that direct bitstream image classification is possible and could be utilised in a first pass database screening of a raw bitstream (within a wired or wireless network) or where computational, memory and bandwidth requirements are severely restricted.
Modern neural network architectures can leverage large amounts of data to generalize well within the training distribution. However, they are less capable of systematic generalization to data drawn from unseen but related distributions, a feat that is hypothesized to require compositional reasoning and reuse of knowledge. In this work, we present Neural Interpreters, an architecture that factorizes inference in a self-attention network as a system of modules, which we call \emph{functions}. Inputs to the model are routed through a sequence of functions in a way that is end-to-end learned. The proposed architecture can flexibly compose computation along width and depth, and lends itself well to capacity extension after training. To demonstrate the versatility of Neural Interpreters, we evaluate it in two distinct settings: image classification and visual abstract reasoning on Raven Progressive Matrices. In the former, we show that Neural Interpreters perform on par with the vision transformer using fewer parameters, while being transferrable to a new task in a sample efficient manner. In the latter, we find that Neural Interpreters are competitive with respect to the state-of-the-art in terms of systematic generalization
Coronavirus disease 2019 (COVID-19) is the highly contagious illness caused by severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2). The standard diagnostic testing procedure for COVID-19 is testing a nasopharyngeal swab for SARS-CoV-2 nucleic acid using a real-time polymerase chain reaction (PCR), which can take multiple days to provide a diagnosis. Another widespread form of testing is rapid antigen testing, which has a low sensitivity compared to PCR, but is favored for its quick diagnosis time of usually 15-30 minutes. Patients who test positive for COVID-19 demonstrate diffuse alveolar damage in 87% of cases. Machine learning has proven to have advantages in image classification problems with radiology. In this work, we introduce CovXR as a machine learning model designed to detect COVID-19 pneumonia in chest X-rays (CXR). CovXR is a convolutional neural network (CNN) trained on over 4,300 chest X-rays. The performance of the model is measured through accuracy, F1 score, sensitivity, and specificity. The model achieves an accuracy of 95.5% and an F1 score of 0.954. The sensitivity is 93.5% and specificity is 97.5%. With accuracy above 95% and F1 score above 0.95, CovXR is highly accurate in predicting COVID-19 pneumonia on CXRs. The model achieves better accuracy than prior work and uses a unique approach to identify COVID-19 pneumonia. CovXR is highly accurate in identifying COVID-19 on CXRs of patients with a PCR confirmed positive diagnosis and provides much faster results than PCR tests.
Ultrasound imaging is a commonly used technology for visualising patient anatomy in real-time during diagnostic and therapeutic procedures. High operator dependency and low reproducibility make ultrasound imaging and interpretation challenging with a steep learning curve. Automatic image classification using deep learning has the potential to overcome some of these challenges by supporting ultrasound training in novices, as well as aiding ultrasound image interpretation in patient with complex pathology for more experienced practitioners. However, the use of deep learning methods requires a large amount of data in order to provide accurate results. Labelling large ultrasound datasets is a challenging task because labels are retrospectively assigned to 2D images without the 3D spatial context available in vivo or that would be inferred while visually tracking structures between frames during the procedure. In this work, we propose a multi-modal convolutional neural network (CNN) architecture that labels endoscopic ultrasound (EUS) images from raw verbal comments provided by a clinician during the procedure. We use a CNN composed of two branches, one for voice data and another for image data, which are joined to predict image labels from the spoken names of anatomical landmarks. The network was trained using recorded verbal comments from expert operators. Our results show a prediction accuracy of 76% at image level on a dataset with 5 different labels. We conclude that the addition of spoken commentaries can increase the performance of ultrasound image classification, and eliminate the burden of manually labelling large EUS datasets necessary for deep learning applications.
The conventional wisdom behind learning deep classification models is to focus on bad-classified examples and ignore well-classified examples that are far from the decision boundary. For instance, when training with cross-entropy loss, examples with higher likelihoods (i.e., well-classified examples) contribute smaller gradients in back-propagation. However, we theoretically show that this common practice hinders representation learning, energy optimization, and margin growth. To counteract this deficiency, we propose to reward well-classified examples with additive bonuses to revive their contribution to the learning process. This counterexample theoretically addresses these three issues. We empirically support this claim by directly verifying the theoretical results or significant performance improvement with our counterexample on diverse tasks, including image classification, graph classification, and machine translation. Furthermore, this paper shows that we can deal with complex scenarios, such as imbalanced classification, OOD detection, and applications under adversarial attacks because our idea can solve these three issues. Code is available at: https://github.com/lancopku/well-classified-examples-are-underestimated.
Domain adaptation techniques have been widely applied to the problem of cross-scene hyperspectral image (HSI) classification. Most existing methods use convolutional neural networks (CNN) to extract statistical features from data and often neglect the potential topological structure information between different land cover classes. CNN-based approaches generally only model the local spatial relationships of the samples, which largely limits their ability to capture the non-local topological relationship that would better represent the underlying data structure of HSI. In order to make up for the above shortcomings, a Topological structure and Semantic information Transfer network (TSTnet) is developed. The method employs the graph structure to characterize topological relationships and the graph convolutional network (GCN) that is good at processing for crossscene HSI classification. In the proposed TSTnet, graph optimal transmission (GOT) is used to align topological relationships to assist distribution alignment between source domain and target domain based on maximum mean difference (MMD). Furthermore, subgraphs from source domain and target domain are dynamically constructed based on CNN features to take advantage of the discriminative capacity of CNN models that in turn improves robustness of classification. In addition, to better characterize the correlation between distribution alignment and topological relationship alignment, a consistency constraint is enforced to integrate the output of CNN and GCN. Experimental results on three cross-scene HSI datasets demonstrate that the proposed TSTnet performs significantly better than some state-of-the-art domain-adaptive approaches.
Recent work suggests that quantum machine learning techniques can be used for classical image classification by encoding the images in quantum states and using a quantum neural network for inference. However, such work has been restricted to very small input images, at most 4 x 4, that are unrealistic and cannot even be accurately labeled by humans. The primary difficulties in using larger input images is that hitherto-proposed encoding schemes necessitate more qubits than are physically realizable. We propose a framework to classify larger, realistic images using quantum systems. Our approach relies on a novel encoding mechanism that embeds images in quantum states while necessitating fewer qubits than prior work. Our framework is able to classify images that are larger than previously possible, up to 16 x 16 for the MNIST dataset on a personal laptop, and obtains accuracy comparable to classical neural networks with the same number of learnable parameters. We also propose a technique for further reducing the number of qubits needed to represent images that may result in an easier physical implementation at the expense of final performance. Our work enables quantum machine learning and classification on classical datasets of dimensions that were previously intractable by physically realizable quantum computers or classical simulation
Most existing neural architecture search (NAS) benchmarks and algorithms prioritize well-studied tasks, e.g. image classification on CIFAR or ImageNet. This makes the performance of NAS approaches in more diverse areas poorly understood. In this paper, we present NAS-Bench-360, a benchmark suite to evaluate methods on domains beyond those traditionally studied in architecture search, and use it to address the following question: do state-of-the-art NAS methods perform well on diverse tasks? To construct the benchmark, we curate ten tasks spanning a diverse array of application domains, dataset sizes, problem dimensionalities, and learning objectives. Each task is carefully chosen to interoperate with modern CNN-based search methods while possibly being far-afield from its original development domain. To speed up and reduce the cost of NAS research, for two of the tasks we release the precomputed performance of 15,625 architectures comprising a standard CNN search space. Experimentally, we show the need for more robust NAS evaluation of the kind NAS-Bench-360 enables by showing that several modern NAS procedures perform inconsistently across the ten tasks, with many catastrophically poor results. We also demonstrate how NAS-Bench-360 and its associated precomputed results will enable future scientific discoveries by testing whether several recent hypotheses promoted in the NAS literature hold on diverse tasks. NAS-Bench-360 is hosted at https://nb360.ml.cmu.edu.
When training and evaluating machine learning models on a large number of tasks, it is important to not only look at average task accuracy -- which may be biased by easy or redundant tasks -- but also worst-case accuracy (i.e. the performance on the task with the lowest accuracy). In this work, we show how to use techniques from the distributionally robust optimization (DRO) literature to improve worst-case performance in multitask learning. We highlight several failure cases of DRO when applied off-the-shelf and present an improved method, Lookahead-DRO (L-DRO), which mitigates these issues. The core idea of L-DRO is to anticipate the interaction between tasks during training in order to choose a dynamic re-weighting of the various task losses, which will (i) lead to minimal worst-case loss and (ii) train on as many tasks as possible. After demonstrating the efficacy of L-DRO on a small controlled synthetic setting, we evaluate it on two realistic benchmarks: a multitask version of the CIFAR-100 image classification dataset and a large-scale multilingual language modeling experiment. Our empirical results show that L-DRO achieves a better trade-off between average and worst-case accuracy with little computational overhead compared to several strong baselines.
We propose ACProp (Asynchronous-centering-Prop), an adaptive optimizer which combines centering of second momentum and asynchronous update (e.g. for $t$-th update, denominator uses information up to step $t-1$, while numerator uses gradient at $t$-th step). ACProp has both strong theoretical properties and empirical performance. With the example by Reddi et al. (2018), we show that asynchronous optimizers (e.g. AdaShift, ACProp) have weaker convergence condition than synchronous optimizers (e.g. Adam, RMSProp, AdaBelief); within asynchronous optimizers, we show that centering of second momentum further weakens the convergence condition. We demonstrate that ACProp has a convergence rate of $O(\frac{1}{\sqrt{T}})$ for the stochastic non-convex case, which matches the oracle rate and outperforms the $O(\frac{logT}{\sqrt{T}})$ rate of RMSProp and Adam. We validate ACProp in extensive empirical studies: ACProp outperforms both SGD and other adaptive optimizers in image classification with CNN, and outperforms well-tuned adaptive optimizers in the training of various GAN models, reinforcement learning and transformers. To sum up, ACProp has good theoretical properties including weak convergence condition and optimal convergence rate, and strong empirical performance including good generalization like SGD and training stability like Adam. We provide the implementation at https://github.com/juntang-zhuang/ACProp-Optimizer.
We study the problem of synthesizing programs that include machine learning components such as deep neural networks (DNNs). We focus on statistical properties, which are properties expected to hold with high probability -- e.g., that an image classification model correctly identifies people in images with high probability. We propose novel algorithms for sketching and synthesizing such programs by leveraging ideas from statistical learning theory to provide statistical soundness guarantees. We evaluate our approach on synthesizing list processing programs that include DNN components used to process image inputs, as well as case studies on image classification and on precision medicine. Our results demonstrate that our approach can be used to synthesize programs with probabilistic guarantees.
Active learning aims to reduce the labeling effort that is required to train algorithms by learning an acquisition function selecting the most relevant data for which a label should be requested from a large unlabeled data pool. Active learning is generally studied on balanced datasets where an equal amount of images per class is available. However, real-world datasets suffer from severe imbalanced classes, the so called long-tail distribution. We argue that this further complicates the active learning process, since the imbalanced data pool can result in suboptimal classifiers. To address this problem in the context of active learning, we proposed a general optimization framework that explicitly takes class-balancing into account. Results on three datasets showed that the method is general (it can be combined with most existing active learning algorithms) and can be effectively applied to boost the performance of both informative and representative-based active learning methods. In addition, we showed that also on balanced datasets our method generally results in a performance gain.
Few-Shot Remote Sensing Scene Classification (FSRSSC) is an important task, which aims to recognize novel scene classes with few examples. Recently, several studies attempt to address the FSRSSC problem by following few-shot natural image classification methods. These existing methods have made promising progress and achieved superior performance. However, they all overlook two unique characteristics of remote sensing images: (i) object co-occurrence that multiple objects tend to appear together in a scene image and (ii) object spatial correlation that these co-occurrence objects are distributed in the scene image following some spatial structure patterns. Such unique characteristics are very beneficial for FSRSSC, which can effectively alleviate the scarcity issue of labeled remote sensing images since they can provide more refined descriptions for each scene class. To fully exploit these characteristics, we propose a novel scene graph matching-based meta-learning framework for FSRSSC, called SGMNet. In this framework, a scene graph construction module is carefully designed to represent each test remote sensing image or each scene class as a scene graph, where the nodes reflect these co-occurrence objects meanwhile the edges capture the spatial correlations between these co-occurrence objects. Then, a scene graph matching module is further developed to evaluate the similarity score between each test remote sensing image and each scene class. Finally, based on the similarity scores, we perform the scene class prediction via a nearest neighbor classifier. We conduct extensive experiments on UCMerced LandUse, WHU19, AID, and NWPU-RESISC45 datasets. The experimental results show that our method obtains superior performance over the previous state-of-the-art methods.