As machine learning (ML) techniques are being increasingly used in many applications, their vulnerability to adversarial attacks becomes well-known. Test time attacks, usually launched by adding adversarial noise to test instances, have been shown effective against the deployed ML models. In practice, one test input may be leveraged by different ML models. Test time attacks targeting a single ML model often neglect their impact on other ML models. In this work, we empirically demonstrate that naively attacking the classifier learning one concept may negatively impact classifiers trained to learn other concepts. For example, for the online image classification scenario, when the Gender classifier is under attack, the (wearing) Glasses classifier is simultaneously attacked with the accuracy dropped from 98.69 to 88.42. This raises an interesting question: is it possible to attack one set of classifiers without impacting the other set that uses the same test instance? Answers to the above research question have interesting implications for protecting privacy against ML model misuse. Attacking ML models that pose unnecessary risks of privacy invasion can be an important tool for protecting individuals from harmful privacy exploitation. In this paper, we address the above research question by developing novel attack techniques that can simultaneously attack one set of ML models while preserving the accuracy of the other. In the case of linear classifiers, we provide a theoretical framework for finding an optimal solution to generate such adversarial examples. Using this theoretical framework, we develop a multi-concept attack strategy in the context of deep learning. Our results demonstrate that our techniques can successfully attack the target classes while protecting the protected classes in many different settings, which is not possible with the existing test-time attack-single strategies.
Current state-of-the-art methods of image classification using convolutional neural networks are often constrained by both latency and power consumption. This places a limit on the devices, particularly low-power edge devices, that can employ these methods. Spiking neural networks (SNNs) are considered to be the third generation of artificial neural networks which aim to address these latency and power constraints by taking inspiration from biological neuronal communication processes. Before data such as images can be input into an SNN, however, they must be first encoded into spike trains. Herein, we propose a method for encoding static images into temporal spike trains using edge detection and an adaptive signal sampling method for use in SNNs. The edge detection process consists of first performing Canny edge detection on the 2D static images and then converting the edge detected images into two X and Y signals using an image-to-signal conversion method. The adaptive signaling approach consists of sampling the signals such that the signals maintain enough detail and are sensitive to abrupt changes in the signal. Temporal encoding mechanisms such as threshold-based representation (TBR) and step-forward (SF) are then able to be used to convert the sampled signals into spike trains. We use various error and indicator metrics to optimize and evaluate the efficiency and precision of the proposed image encoding approach. Comparison results between the original and reconstructed signals from spike trains generated using edge-detection and adaptive temporal encoding mechanism exhibit 18x and 7x reduction in average root mean square error (RMSE) compared to the conventional SF and TBR encoding, respectively, while used for encoding MNIST dataset.
Group Convolutional Neural Networks (G-CNNs) constrain learned features to respect the symmetries in the selected group, and lead to better generalization when these symmetries appear in the data. If this is not the case, however, equivariance leads to overly constrained models and worse performance. Frequently, transformations occurring in data can be better represented by a subset of a group than by a group as a whole, e.g., rotations in $[-90^{\circ}, 90^{\circ}]$. In such cases, a model that respects equivariance $\textit{partially}$ is better suited to represent the data. In addition, relevant transformations may differ for low and high-level features. For instance, full rotation equivariance is useful to describe edge orientations in a face, but partial rotation equivariance is better suited to describe face poses relative to the camera. In other words, the optimal level of equivariance may differ per layer. In this work, we introduce $\textit{Partial G-CNNs}$: G-CNNs able to learn layer-wise levels of partial and full equivariance to discrete, continuous groups and combinations thereof as part of training. Partial G-CNNs retain full equivariance when beneficial, e.g., for rotated MNIST, but adjust it whenever it becomes harmful, e.g., for classification of 6 / 9 digits or natural images. We empirically show that partial G-CNNs pair G-CNNs when full equivariance is advantageous, and outperform them otherwise.
In vision domain, large-scale natural datasets typically exhibit long-tailed distribution which has large class imbalance between head and tail classes. This distribution poses difficulty in learning good representations for tail classes. Recent developments have shown good long-tailed model can be learnt by decoupling the training into representation learning and classifier balancing. However, these works pay insufficient consideration on the long-tailed effect on representation learning. In this work, we propose interpolative centroid contrastive learning (ICCL) to improve long-tailed representation learning. ICCL interpolates two images from a class-agnostic sampler and a class-aware sampler, and trains the model such that the representation of the interpolative image can be used to retrieve the centroids for both source classes. We demonstrate the effectiveness of our approach on multiple long-tailed image classification benchmarks. Our result shows a significant accuracy gain of 2.8% on the iNaturalist 2018 dataset with a real-world long-tailed distribution.
With ever growing scale of neural models, knowledge distillation (KD) attracts more attention as a prominent tool for neural model compression. However, there are counter intuitive observations in the literature showing some challenging limitations of KD. A case in point is that the best performing checkpoint of the teacher might not necessarily be the best teacher for training the student in KD. Therefore, one important question would be how to find the best checkpoint of the teacher for distillation? Searching through the checkpoints of the teacher would be a very tedious and computationally expensive process, which we refer to as the \textit{checkpoint-search problem}. Moreover, another observation is that larger teachers might not necessarily be better teachers in KD which is referred to as the \textit{capacity-gap} problem. To address these challenging problems, in this work, we introduce our progressive knowledge distillation (Pro-KD) technique which defines a smoother training path for the student by following the training footprints of the teacher instead of solely relying on distilling from a single mature fully-trained teacher. We demonstrate that our technique is quite effective in mitigating the capacity-gap problem and the checkpoint search problem. We evaluate our technique using a comprehensive set of experiments on different tasks such as image classification (CIFAR-10 and CIFAR-100), natural language understanding tasks of the GLUE benchmark, and question answering (SQuAD 1.1 and 2.0) using BERT-based models and consistently got superior results over state-of-the-art techniques.
Deep neural networks (DNNs) are now the de facto choice for computer vision tasks such as image classification. However, their complexity and "black box" nature often renders the systems they're deployed in vulnerable to a range of security threats. Successfully identifying such threats, especially in safety-critical real-world applications is thus of utmost importance, but still very much an open problem. We present TESDA, a low-overhead, flexible, and statistically grounded method for {online detection} of attacks by exploiting the discrepancies they cause in the distributions of intermediate layer features of DNNs. Unlike most prior work, we require neither dedicated hardware to run in real-time, nor the presence of a Trojan trigger to detect discrepancies in behavior. We empirically establish our method's usefulness and practicality across multiple architectures, datasets and diverse attacks, consistently achieving detection coverages of above 95% with operation count overheads as low as 1-2%.
Machine learning in medical research, by nature, needs careful attention on obeying the regulations of data privacy, making it difficult to train a machine learning model over gathered data from different medical centers. Failure of leveraging data of the same kind may result in poor generalizability for the trained model. Federated learning (FL) enables collaboratively training a joint model while keeping the data decentralized for multiple medical centers. However, federated optimizations often suffer from the heterogeneity of the data distribution across medical centers. In this work, we propose Federated Learning with Shared Label Distribution (FedSLD) for classification tasks, a method that assumes knowledge of the label distributions for all the participating clients in the federation. FedSLD adjusts the contribution of each data sample to the local objective during optimization given knowledge of the distribution, mitigating the instability brought by data heterogeneity across all clients. We conduct extensive experiments on four publicly available image datasets with different types of non-IID data distributions. Our results show that FedSLD achieves better convergence performance than the compared leading FL optimization algorithms, increasing the test accuracy by up to 5.50 percentage points.
Distributed data-parallel training has been widely adopted for deep neural network (DNN) models. Although current deep learning (DL) frameworks scale well for dense models like image classification models, we find that these DL frameworks have relatively low scalability for sparse models like natural language processing (NLP) models that have highly sparse embedding tables. Most existing works overlook the sparsity of model parameters thus suffering from significant but unnecessary communication overhead. In this paper, we propose EmbRace, an efficient communication framework to accelerate communications of distributed training for sparse models. EmbRace introduces Sparsity-aware Hybrid Communication, which integrates AlltoAll and model parallelism into data-parallel training, so as to reduce the communication overhead of highly sparse parameters. To effectively overlap sparse communication with both backward and forward computation, EmbRace further designs a 2D Communication Scheduling approach which optimizes the model computation procedure, relaxes the dependency of embeddings, and schedules the sparse communications of each embedding row with a priority queue. We have implemented a prototype of EmbRace based on PyTorch and Horovod, and conducted comprehensive evaluations with four representative NLP models. Experimental results show that EmbRace achieves up to 2.41X speedup compared to the state-of-the-art distributed training baselines.
Heavy ball momentum is crucial in accelerating (stochastic) gradient-based optimization algorithms for machine learning. Existing heavy ball momentum is usually weighted by a uniform hyperparameter, which relies on excessive tuning. Moreover, the calibrated fixed hyperparameter may not lead to optimal performance. In this paper, to eliminate the effort for tuning the momentum-related hyperparameter, we propose a new adaptive momentum inspired by the optimal choice of the heavy ball momentum for quadratic optimization. Our proposed adaptive heavy ball momentum can improve stochastic gradient descent (SGD) and Adam. SGD and Adam with the newly designed adaptive momentum are more robust to large learning rates, converge faster, and generalize better than the baselines. We verify the efficiency of SGD and Adam with the new adaptive momentum on extensive machine learning benchmarks, including image classification, language modeling, and machine translation. Finally, we provide convergence guarantees for SGD and Adam with the proposed adaptive momentum.
Contrastive learning is a powerful technique to learn representations that are semantically distinctive and geometrically invariant. While most of the earlier approaches have demonstrated its effectiveness on single-modality learning tasks such as image classification, recently there have been a few attempts towards extending this idea to multi-modal data. In this paper, we propose two loss functions based on normalized cross-entropy to perform the task of learning joint visual-semantic embedding using batch contrastive training. In a batch, for a given anchor point from one modality, we consider its negatives only from another modality, and define our first contrastive loss based on expected violations incurred by all the negatives. Next, we update this loss and define the second contrastive loss based on the violation incurred only by the hardest negative. We compare our results with existing visual-semantic embedding methods on cross-modal image-to-text and text-to-image retrieval tasks using the MS-COCO and Flickr30K datasets, where we outperform the state-of-the-art on the MS-COCO dataset and achieve comparable results on the Flickr30K dataset.
Deep-learning-based image classification frameworks often suffer from the noisy label problem caused by the inter-observer variation. Recent studies employed learning-to-learn paradigms (e.g., Co-teaching and JoCoR) to filter the samples with noisy labels from the training set. However, most of them use a simple cross-entropy loss as the criterion for noisy label identification. The hard samples, which are beneficial for classifier learning, are often mistakenly treated as noises in such a setting since both the hard samples and ones with noisy labels lead to a relatively larger loss value than the easy cases. In this paper, we propose a plugin module, namely noise ignoring block (NIB), consisting of a probability transition matrix and an inter-class correlation (IC) loss, to separate the hard samples from the mislabeled ones, and further boost the accuracy of image classification network trained with noisy labels. Concretely, our IC loss is calculated as Kullback-Leibler divergence between the network prediction and the accumulative soft label generated by the probability transition matrix. Such that, with the lower value of IC loss, the hard cases can be easily distinguished from mislabeled cases. Extensive experiments are conducted on natural and medical image datasets (CIFAR-10 and ISIC 2019). The experimental results show that our NIB module consistently improves the performances of the state-of-the-art robust training methods.
Continual learning in online scenario aims to learn a sequence of new tasks from data stream using each data only once for training, which is more realistic than in offline mode assuming data from new task are all available. However, this problem is still under-explored for the challenging class-incremental setting in which the model classifies all classes seen so far during inference. Particularly, performance struggles with increased number of tasks or additional classes to learn for each task. In addition, most existing methods require storing original data as exemplars for knowledge replay, which may not be feasible for certain applications with limited memory budget or privacy concerns. In this work, we introduce an effective and memory-efficient method for online continual learning under class-incremental setting through candidates selection from each learned task together with prior incorporation using stored feature embeddings instead of original data as exemplars. Our proposed method implemented for image classification task achieves the best results under different benchmark datasets for online continual learning including CIFAR-10, CIFAR-100 and CORE-50 while requiring much less memory resource compared with existing works.
Network pruning reduces the size of neural networks by removing (pruning) neurons such that the performance drop is minimal. Traditional pruning approaches focus on designing metrics to quantify the usefulness of a neuron which is often quite tedious and sub-optimal. More recent approaches have instead focused on training auxiliary networks to automatically learn how useful each neuron is however, they often do not take computational limitations into account. In this work, we propose a general methodology for pruning neural networks. Our proposed methodology can prune neural networks to respect pre-defined computational budgets on arbitrary, possibly non-differentiable, functions. Furthermore, we only assume the ability to be able to evaluate these functions for different inputs, and hence they do not need to be fully specified beforehand. We achieve this by proposing a novel pruning strategy via constrained reinforcement learning algorithms. We prove the effectiveness of our approach via comparison with state-of-the-art methods on standard image classification datasets. Specifically, we reduce 83-92.90 of total parameters on various variants of VGG while achieving comparable or better performance than that of original networks. We also achieved 75.09 reduction in parameters on ResNet18 without incurring any loss in accuracy.
The application of machine learning techniques to the medical domain is especially challenging due to the required level of precision and the incurrence of huge risks of minute errors. Employing these techniques to a more complex subdomain of hematological diagnosis seems quite promising, with automatic identification of blood cell types, which can help in detection of hematologic disorders. In this paper, we benchmark 27 popular deep convolutional neural network architectures on the microscopic peripheral blood cell images dataset. The dataset is publicly available, with large number of normal peripheral blood cells acquired using the CellaVision DM96 analyzer and identified by expert pathologists into eight different cell types. We fine-tune the state-of-the-art image classification models pre-trained on the ImageNet dataset for blood cell classification. We exploit data augmentation techniques during training to avoid overfitting and achieve generalization. An ensemble of the top performing models obtains significant improvements over past published works, achieving the state-of-the-art results with a classification accuracy of 99.51%. Our work provides empirical baselines and benchmarks on standard deep-learning architectures for microscopic peripheral blood cell recognition task.
We present a High-Resolution Transformer (HRFormer) that learns high-resolution representations for dense prediction tasks, in contrast to the original Vision Transformer that produces low-resolution representations and has high memory and computational cost. We take advantage of the multi-resolution parallel design introduced in high-resolution convolutional networks (HRNet), along with local-window self-attention that performs self-attention over small non-overlapping image windows, for improving the memory and computation efficiency. In addition, we introduce a convolution into the FFN to exchange information across the disconnected image windows. We demonstrate the effectiveness of the High-Resolution Transformer on both human pose estimation and semantic segmentation tasks, e.g., HRFormer outperforms Swin transformer by $1.3$ AP on COCO pose estimation with $50\%$ fewer parameters and $30\%$ fewer FLOPs. Code is available at: https://github.com/HRNet/HRFormer.
We introduce Network Augmentation (NetAug), a new training method for improving the performance of tiny neural networks. Existing regularization techniques (e.g., data augmentation, dropout) have shown much success on large neural networks by adding noise to overcome over-fitting. However, we found these techniques hurt the performance of tiny neural networks. We argue that training tiny models are different from large models: rather than augmenting the data, we should augment the model, since tiny models tend to suffer from under-fitting rather than over-fitting due to limited capacity. To alleviate this issue, NetAug augments the network (reverse dropout) instead of inserting noise into the dataset or the network. It puts the tiny model into larger models and encourages it to work as a sub-model of larger models to get extra supervision, in addition to functioning as an independent model. At test time, only the tiny model is used for inference, incurring zero inference overhead. We demonstrate the effectiveness of NetAug on image classification and object detection. NetAug consistently improves the performance of tiny models, achieving up to 2.2% accuracy improvement on ImageNet. On object detection, achieving the same level of performance, NetAug requires 41% fewer MACs on Pascal VOC and 38% fewer MACs on COCO than the baseline.
Training from scratch is the most common way to build a Convolutional Neural Network (CNN) based model. What if we can build new CNN models by reusing parts from previously build CNN models? What if we can improve a CNN model by replacing (possibly faulty) parts with other parts? In both cases, instead of training, can we identify the part responsible for each output class (module) in the model(s) and reuse or replace only the desired output classes to build a model? Prior work has proposed decomposing dense-based networks into modules (one for each output class) to enable reusability and replaceability in various scenarios. However, this work is limited to the dense layers and based on the one-to-one relationship between the nodes in consecutive layers. Due to the shared architecture in the CNN model, prior work cannot be adapted directly. In this paper, we propose to decompose a CNN model used for image classification problems into modules for each output class. These modules can further be reused or replaced to build a new model. We have evaluated our approach with CIFAR-10, CIFAR-100, and ImageNet tiny datasets with three variations of ResNet models and found that enabling decomposition comes with a small cost (2.38% and 0.81% for top-1 and top-5 accuracy, respectively). Also, building a model by reusing or replacing modules can be done with a 2.3% and 0.5% average loss of accuracy. Furthermore, reusing and replacing these modules reduces CO2e emission by ~37 times compared to training the model from scratch.
Existing transfer attack methods commonly assume that the attacker knows the training set (e.g., the label set, the input size) of the black-box victim models, which is usually unrealistic because in some cases the attacker cannot know this information. In this paper, we define a Generalized Transferable Attack (GTA) problem where the attacker doesn't know this information and is acquired to attack any randomly encountered images that may come from unknown datasets. To solve the GTA problem, we propose a novel Image Classification Eraser (ICE) that trains a particular attacker to erase classification information of any images from arbitrary datasets. Experiments on several datasets demonstrate that ICE greatly outperforms existing transfer attacks on GTA, and show that ICE uses similar texture-like noises to perturb different images from different datasets. Moreover, fast fourier transformation analysis indicates that the main components in each ICE noise are three sine waves for the R, G, and B image channels. Inspired by this interesting finding, we then design a novel Sine Attack (SA) method to optimize the three sine waves. Experiments show that SA performs comparably to ICE, indicating that the three sine waves are effective and enough to break DNNs under the GTA setting.
As datasets and models become increasingly large, distributed training has become a necessary component to allow deep neural networks to train in reasonable amounts of time. However, distributed training can have substantial communication overhead that hinders its scalability. One strategy for reducing this overhead is to perform multiple unsynchronized SGD steps independently on each worker between synchronization steps, a technique known as local SGD. We conduct a comprehensive empirical study of local SGD and related methods on a large-scale image classification task. We find that performing local SGD comes at a price: lower communication costs (and thereby faster training) are accompanied by lower accuracy. This finding is in contrast from the smaller-scale experiments in prior work, suggesting that local SGD encounters challenges at scale. We further show that incorporating the slow momentum framework of Wang et al. (2020) consistently improves accuracy without requiring additional communication, hinting at future directions for potentially escaping this trade-off.
Vacuum Insulated Glazing (VIG) is a highly thermally insulating window technology, which boasts an extremely thin profile and lower weight as compared to gas-filled insulated glazing units of equivalent performance. The VIG is a double-pane configuration with a submillimeter vacuum gap between the panes and therefore under constant atmospheric pressure over their service life. Small pillars are positioned between the panes to maintain the gap, which can damage the glass reducing the lifetime of the VIG unit. To efficiently assess any surface damage on the glass, an automated damage detection system is highly desirable. For the purpose of classifying the damage, we have developed, trained, and tested a deep learning computer vision system using convolutional neural networks. The classification model flawlessly classified the test dataset with an area under the curve (AUC) for the receiver operating characteristic (ROC) of 100%. We have automatically cropped the images down to their relevant information by using Faster-RCNN to locate the position of the pillars. We employ the state-of-the-art methods Grad-CAM and Score-CAM of explainable Artificial Intelligence (XAI) to provide an understanding of the internal mechanisms and were able to show that our classifier outperforms ResNet50V2 for identification of crack locations and geometry. The proposed methods can therefore be used to detect systematic defects even without large amounts of training data. Further analyses of our model's predictive capabilities demonstrates its superiority over state-of-the-art models (ResNet50V2, ResNet101V2 and ResNet152V2) in terms of convergence speed, accuracy, precision at 100% recall and AUC for ROC.