Knowledge distillation(KD) is a widely-used technique to train compact models in object detection. However, there is still a lack of study on how to distill between heterogeneous detectors. In this paper, we empirically find that better FPN features from a heterogeneous teacher detector can help the student although their detection heads and label assignments are different. However, directly aligning the feature maps to distill detectors suffers from two problems. First, the difference in feature magnitude between the teacher and the student could enforce overly strict constraints on the student. Second, the FPN stages and channels with large feature magnitude from the teacher model could dominate the gradient of distillation loss, which will overwhelm the effects of other features in KD and introduce much noise. To address the above issues, we propose to imitate features with Pearson Correlation Coefficient to focus on the relational information from the teacher and relax constraints on the magnitude of the features. Our method consistently outperforms the existing detection KD methods and works for both homogeneous and heterogeneous student-teacher pairs. Furthermore, it converges faster. With a powerful MaskRCNN-Swin detector as the teacher, ResNet-50 based RetinaNet and FCOS achieve 41.5% and 43.9% mAP on COCO2017, which are 4.1\% and 4.8\% higher than the baseline, respectively.

Early evaluation of patients who require special care and who have high death-expectancy in COVID-19, and the effective determination of relevant biomarkers on large sample-groups are important to reduce mortality. This study aimed to reveal the routine blood-value predictors of COVID-19 mortality and to determine the lethal-risk levels of these predictors during the disease process. The dataset of the study consists of 38 routine blood-values of 2597 patients who died (n = 233) and those who recovered (n = 2364) from COVID-19 in August-December, 2021. In this study, the histogram-based gradient-boosting (HGB) model was the most successful machine-learning classifier in detecting living and deceased COVID-19 patients (with squared F1 metrics F1^2 = 1). The most efficient binary combinations with procalcitonin were obtained with D-dimer, ESR, D-Bil and ferritin. The HGB model operated with these feature pairs correctly detected almost all of the patients who survived and those who died (precision > 0.98, recall > 0.98, F1^2 > 0.98). Furthermore, in the HGB model operated with a single feature, the most efficient features were procalcitonin (F1^2 = 0.96) and ferritin (F1^2 = 0.91). In addition, according to the two-threshold approach, ferritin values between 376.2 mkg/L and 396.0 mkg/L (F1^2 = 0.91) and pro-calcitonin values between 0.2 mkg/L and 5.2 mkg/L (F1^2 = 0.95) were found to be fatal risk levels for COVID-19. Considering all the results, we suggest that many features combined with these features, especially procalcitonin and ferritin, operated with the HGB model, can be used to achieve very successful results in the classification of those who live, and those who die from COVID-19. Moreover, we strongly recommend that clinicians consider the critical levels we have found for procalcitonin and ferritin properties, to reduce the lethality of the COVID-19 disease.

The NASA Astrophysics Data System (ADS) is an essential tool for researchers that allows them to explore the astronomy and astrophysics scientific literature, but it has yet to exploit recent advances in natural language processing. At ADASS 2021, we introduced astroBERT, a machine learning language model tailored to the text used in astronomy papers in ADS. In this work we: 1. announce the first public release of the astroBERT language model; 2. show how astroBERT improves over existing public language models on astrophysics specific tasks; 3. and detail how ADS plans to harness the unique structure of scientific papers, the citation graph and citation context, to further improve astroBERT.

The recent success and proliferation of machine learning and deep learning have provided powerful tools, which are also utilized for encrypted traffic analysis, classification, and threat detection in computer networks. These methods, neural networks in particular, are often complex and require a huge corpus of training data. Therefore, this paper focuses on collecting a large up-to-date dataset with almost 200 fine-grained service labels and 140 million network flows extended with packet-level metadata. The number of flows is three orders of magnitude higher than in other existing public labeled datasets of encrypted traffic. The number of service labels, which is important to make the problem hard and realistic, is four times higher than in the public dataset with the most class labels. The published dataset is intended as a benchmark for identifying services in encrypted traffic. Service identification can be further extended with the task of "rejecting" unknown services, i.e., the traffic not seen during the training phase. Neural networks offer superior performance for tackling this more challenging problem. To showcase the dataset's usefulness, we implemented a neural network with a multi-modal architecture, which is the state-of-the-art approach, and achieved 97.04% classification accuracy and detected 91.94% of unknown services with 5% false positive rate.

Unlike the typical classification setting where each instance is associated with a single class, in multi-label learning each instance is associated with multiple classes simultaneously. Therefore the learning task in this setting is to predict the subset of classes to which each instance belongs. This work examines the application of a recently developed framework called Conformal Prediction (CP) to the multi-label learning setting. CP complements the predictions of machine learning algorithms with reliable measures of confidence. As a result the proposed approach instead of just predicting the most likely subset of classes for a new unseen instance, also indicates the likelihood of each predicted subset being correct. This additional information is especially valuable in the multi-label setting where the overall uncertainty is extremely high.

How do we know when the predictions made by a classifier can be trusted? This is a fundamental problem that also has immense practical applicability, especially in safety-critical areas such as medicine and autonomous driving. The de facto approach of using the classifier's softmax outputs as a proxy for trustworthiness suffers from the over-confidence issue; while the most recent works incur problems such as additional retraining cost and accuracy versus trustworthiness trade-off. In this work, we argue that the trustworthiness of a classifier's prediction for a sample is highly associated with two factors: the sample's neighborhood information and the classifier's output. To combine the best of both worlds, we design a model-agnostic post-hoc approach NeighborAgg to leverage the two essential information via an adaptive neighborhood aggregation. Theoretically, we show that NeighborAgg is a generalized version of a one-hop graph convolutional network, inheriting the powerful modeling ability to capture the varying similarity between samples within each class. We also extend our approach to the closely related task of mislabel detection and provide a theoretical coverage guarantee to bound the false negative. Empirically, extensive experiments on image and tabular benchmarks verify our theory and suggest that NeighborAgg outperforms other methods, achieving state-of-the-art trustworthiness performance.

Entity Alignment (EA), which aims to detect entity mappings (i.e. equivalent entity pairs) in different Knowledge Graphs (KGs), is critical for KG fusion. Neural EA methods dominate current EA research but still suffer from their reliance on labelled mappings. To solve this problem, a few works have explored boosting the training of EA models with self-training, which adds confidently predicted mappings into the training data iteratively. Though the effectiveness of self-training can be glimpsed in some specific settings, we still have very limited knowledge about it. One reason is the existing works concentrate on devising EA models and only treat self-training as an auxiliary tool. To fill this knowledge gap, we change the perspective to self-training to shed light on it. In addition, the existing self-training strategies have limited impact because they introduce either much False Positive noise or a low quantity of True Positive pseudo mappings. To improve self-training for EA, we propose exploiting the dependencies between entities, a particularity of EA, to suppress the noise without hurting the recall of True Positive mappings. Through extensive experiments, we show that the introduction of dependency makes the self-training strategy for EA reach a new level. The value of self-training in alleviating the reliance on annotation is actually much higher than what has been realised. Furthermore, we suggest future study on smart data annotation to break the ceiling of EA performance.

In recent years, applying deep learning (DL) to assess structural damages has gained growing popularity in vision-based structural health monitoring (SHM). However, both data deficiency and class-imbalance hinder the wide adoption of DL in practical applications of SHM. Common mitigation strategies include transfer learning, over-sampling, and under-sampling, yet these ad-hoc methods only provide limited performance boost that varies from one case to another. In this work, we introduce one variant of the Generative Adversarial Network (GAN), named the balanced semi-supervised GAN (BSS-GAN). It adopts the semi-supervised learning concept and applies balanced-batch sampling in training to resolve low-data and imbalanced-class problems. A series of computer experiments on concrete cracking and spalling classification were conducted under the low-data imbalanced-class regime with limited computing power. The results show that the BSS-GAN is able to achieve better damage detection in terms of recall and $F_\beta$ score than other conventional methods, indicating its state-of-the-art performance.

We aim for image-based novelty detection. Despite considerable progress, existing models either fail or face a dramatic drop under the so-called "near-distribution" setting, where the differences between normal and anomalous samples are subtle. We first demonstrate existing methods experience up to 20% decrease in performance in the near-distribution setting. Next, we propose to exploit a score-based generative model to produce synthetic near-distribution anomalous data. Our model is then fine-tuned to distinguish such data from the normal samples. We provide a quantitative as well as qualitative evaluation of this strategy, and compare the results with a variety of GAN-based models. Effectiveness of our method for both the near-distribution and standard novelty detection is assessed through extensive experiments on datasets in diverse applications such as medical images, object classification, and quality control. This reveals that our method considerably improves over existing models, and consistently decreases the gap between the near-distribution and standard novelty detection performance. Such samples are called anomalous, while the training set is referred to as normal. One has access to only normal data during training in ND. Recently, PANDA (33) and CSI (43) have considerably pushed state-of-the-art and achieved more than 90% the area under the receiver operating characteristics (AUROC) on the CIFAR-10 dataset (22) in the ND task, where one class is assumed to be normal and the rest are considered anomalous. However, as we will show empirically, these methods struggle to achieve a similar performance in situations where outliers are semantically close to the normal distribution, e.g. In the literature novelty detection and anomaly detection are used interchangeably.

Learned classifiers should often possess certain invariance properties meant to encourage fairness, robustness, or out-of-distribution generalization. However, multiple recent works empirically demonstrate that common invariance-inducing regularizers are ineffective in the over-parameterized regime, in which classifiers perfectly fit (i.e. interpolate) the training data. This suggests that the phenomenon of ``benign overfitting," in which models generalize well despite interpolating, might not favorably extend to settings in which robustness or fairness are desirable. In this work we provide a theoretical justification for these observations. We prove that -- even in the simplest of settings -- any interpolating learning rule (with arbitrarily small margin) will not satisfy these invariance properties. We then propose and analyze an algorithm that -- in the same setting -- successfully learns a non-interpolating classifier that is provably invariant. We validate our theoretical observations on simulated data and the Waterbirds dataset.