Relieving the Incompatibility of Network Representation and Classification for Long-Tailed Data Distribution

In the real-world scenario, data often have a long-tailed distribution and training deep neural networks on such an imbalanced dataset has become a great challenge. The main problem caused by a long-tailed data distribution is that common classes will dominate the training results and achieve a very low accuracy on the rare classes. Recent work focuses on improving the network representation ability to overcome the long-tailed problem, while it always ignores adapting the network classifier to a long-tailed case, which will cause the “incompatibility” problem of network representation and network classifier. In this paper, we use knowledge distillation to solve the long-tailed data distribution problem and fully optimize the network representation and classifier simultaneously. We propose multiexperts knowledge distillation with class-balanced sampling to jointly learn high-quality network representation and classifier. Also, a channel activation-based knowledge distillation method is also proposed to improve the performance further. State-of-the-art performance on several large-scale long-tailed classification datasets shows the superior generalization of our method.

Using Progressive Context Encoders for Anomaly Detection in Digital Pathology Images

Whole slide imaging (WSI) is transforming the practice of pathology, converting a qualitative discipline into a quantitative one. However, one must exercise caution in interpreting algorithm assertions, particularly in pathology where an incorrect classification could have profound impacts on a patient, and rare classes exist that may not have been seen by the algorithm during training. A more robust approach would be to identify areas of an image for which the pathologist should concentrate their effort to make a final diagnosis. This anomaly detection strategy would be ideal for WSI, but given the extremely high resolution and large file sizes, such an approach is difficult. Here, we combine progressive generative adversarial networks with a flexible adversarial autoencoder architecture capable of learning the normal distribution of WSIs of normal skin tissue at extremely high resolution and demonstrate its anomaly detection performance. Our approach yielded pixel-level accuracy of 89% for identifying melanoma, suggesting that our label-free anomaly detection pipeline is a viable strategy for generating high quality annotations - without tedious manual segmentation by pathologists. The code is publicly available at https://github.com/Steven-N-Hart/P-CEAD.

Rare Traffic Signs Recognition

Recognition of road signs is an important part of the control systems of autonomous vehicles and driver assistance systems. Modern recognition methods based on neural networks require large well-labeled datasets. Marking up data is quite expensive, but it is even more difficult to mark up rare classes of objects. To solve this problem in this article, we use synthetic data. We improve the marking of the Russian traffic signs dataset (RTSD) in semi-automatic mode adding 9 thousand new road signs. We perform an experimental evaluation of the currently best classifiers and detectors in the task of recognizing road signs. To improve the performance of classification, we use stochastic weight averaging (SWA) and contrastive loss. The use of modern methods allows us to train a high-quality neural network on synthetic data, which was previously impossible, and significantly improves the metrics of recognition of both rare and frequent road signs.

Ground-based visible spectroscopy of asteroids to support the development of an unsupervised Gaia asteroid taxonomy

Context. The Gaia mission of the European Space Agency is measuring reflectance spectra of a number to the order of 105 small Solar System objects. A first sample will be published in the Gaia Data Release scheduled for 2021. Aims. The aim of our work was to test the procedure developed to obtain taxonomic classifications for asteroids based only on Gaia spectroscopic data. Methods. We used asteroid spectra obtained using the DOLORES (Device Optimised for the LOw RESolution) instrument, a low-resolution spectrograph and camera installed at the Nasmyth B focus of the Telescopio Nazionale Galileo. Because these spectra have a higher spectral resolution than that typical of the Gaia spectra, we resampled them to more closely match the expected Gaia spectral resolution. We then developed a cloning algorithm to build a database of asteroid spectra belonging to a variety of taxonomic classes, starting from a set of 33 prototypes chosen from the 50 asteroids in our observing campaign. We used them to generate a simulated population of 10 000 representative asteroid spectra and employed them as the input to the algorithm for taxonomic classification developed to analyze Gaia asteroid spectra. Results. Using the simulated population of 10 000 representative asteroid spectra in the algorithm to be used to produce the Gaia asteroid taxonomy at the end of the mission, we found 12 distinct taxonomic classes. Two of them, with 53% of the sample, are dominant. At the other extreme are three classes each with <1% of the sample, and these consist of the previously known rare classes A, D/Ld, and V; 99.1% of the simulated population fall into a single class. Conclusions. We demonstrated the robustness of our algorithm for taxonomic classification by using a sample of simulated asteroid spectra fully representative of what is expected to be in the Gaia spectroscopic data catalogue for asteroids. Increasingly larger data sets will become available as soon as they are published in the future Gaia data releases, with the next one coming in 2021. This will be exploited to develop a correspondingly improved taxonomy, likely with minor tweaks to the algorithm described here, as suggested by the results of this preliminary analysis.

A Deep Neural Network to Distinguish COVID-19 from other Chest Diseases using X-ray Images

Background: Scanning patient’s lungs to detect a Coronavirus 2019 (COVID-19) may lead to similar imaging with other chest diseases that strongly requires a multidisciplinary approach to confirm the diagnosis. There are only few works targeted pathological x-ray images. Most of the works targeted only single disease detection which is not good enough. Some works have provided for all classes however the results suffer due to lack of data for rare classes and data unbalancing problem. Methods: Due to arise of COVID-19 virus medical facilities of many countries are overwhelmed and there is a need of intelligent system to detect it. There have been few works regarding detection of the coronavirus but there are many cases where it can be misclassified as some techniques do not provide any goodness if it can only identify type of diseases and ignore the rest. This work is a deep learning-based model to distinguish between cases of COVID-19 from other chest diseases which is need of today. Results: A Deep Neural Network model provides a significant contribution in terms of detecting COVID-19 and provide effective analysis of chest related diseases with respect to age and gender. Our model achieves 87% accuracy in terms of Gan based synthetic data and four different types of deep learning- based models which provided state of the art comparable results. Conclusion: If the gap in identifying of all viral pneumonias is not filled with effective automation of chest disease detection the healthcare industry may have to bear unfavorable circumstances.

Tell Me What They're Holding: Weakly-Supervised Object Detection with Transferable Knowledge from Human-Object Interaction

In this work, we introduce a novel weakly supervised object detection (WSOD) paradigm to detect objects belonging to rare classes that have not many examples using transferable knowledge from human-object interactions (HOI). While WSOD shows lower performance than full supervision, we mainly focus on HOI as the main context which can strongly supervise complex semantics in images. Therefore, we propose a novel module called RRPN (relational region proposal network) which outputs an object-localizing attention map only with human poses and action verbs. In the source domain, we fully train an object detector and the RRPN with full supervision of HOI. With transferred knowledge about localization map from the trained RRPN, a new object detector can learn unseen objects with weak verbal supervision of HOI without bounding box annotations in the target domain. Because the RRPN is designed as an add-on type, we can apply it not only to the object detection but also to other domains such as semantic segmentation. The experimental results on HICO-DET dataset show the possibility that the proposed method can be a cheap alternative for the current supervised object detection paradigm. Moreover, qualitative results demonstrate that our model can properly localize unseen objects on HICO-DET and V-COCO datasets.

Synthesis of substituted benzo[b][1,4]oxazepine derivatives by the reaction of 2-aminophenols with alkynones

We have developed a novel synthetic method accessing benzo[b][1,4]oxazepines that are one of the rare classes of benzoxazepine derivatives by reaction of 2-aminophenols with alkynones in 1,4-dioxane at 100 °C.