No abstract

Convolutional neural networks have gained a remarkable success in computer vision. However, most usable network architectures are hand-crafted and usually require expertise and elaborate design. In this paper, we provide a block-wise network generation pipeline called BlockQNN which automatically builds high-performance networks using the Q-Learning paradigm with epsilon-greedy exploration strategy. The optimal network block is constructed by the learning agent which is trained sequentially to choose component layers. We stack the block to construct the whole auto-generated network. To accelerate the generation process, we also propose a distributed asynchronous framework and an early stop strategy. The block-wise generation brings unique advantages: (1) it performs competitive results in comparison to the hand-crafted state-of-the-art networks on image classification, additionally, the best network generated by BlockQNN achieves 3.54% top-1 error rate on CIFAR-10 which beats all existing auto-generate networks.(2) in the meanwhile, it offers tremendous reduction of the search space in designing networks which only spends 3 days with 32 GPUs, and (3) moreover, it has strong generalizability that the network built on CIFAR also performs well on a larger-scale ImageNet dataset. * This work was done when Zhao Zhong worked as an intern at Sense-Time Research. performance network architecture generally possesses a tremendous number of possible configurations about the number of layers, hyperparameters in each layer and type of each layer. It is hence infeasible for manually exhaustive search, and the design of successful hand-crafted networks heavily rely on expert knowledge and experience. Therefore, constructing the network in a smart and automatic manner remains an open problem.Although some recent works have attempted computeraided or automated network design [2,37], there are several challenges still unsolved: (1) Modern neural networks always consist of hundreds of convolutional layers, each of which has numerous options in type and hyperparameters. It makes a huge search space and heavy computational costs for network generation. (2) One typically designed network is usually limited on a specific dataset or task, and thus is hard to transfer to other tasks or generalize to another dataset with different input data sizes. In this paper, we provide a solution to the aforementioned challenges by a novel fast Q-learning framework, called BlockQNN, to automatically design the network architecture, as shown in Fig. 1.Particularly, to make the network generation efficient and generalizable, we introduce the block-wise network generation, i.e., we construct the network architecture as a flexible stack of personalized blocks rather tedious per-layer network piling. Indeed, most modern CNN architectures such as Inception [30,14,31] and ResNet Series [10,11] are assembled as the stack of basic block structures. For example, the inception and residual blocks shown in Fig. 1 are repeatedly concatenated to construct the enti...

Background Previous studies have identified asthma phenotypes based on small numbers of clinical, physiologic or inflammatory characteristics. However, no studies have utilized a wide range of variables using machine learning approaches. Objectives To identify subphenotypes of asthma utilizing blood, bronchoscopic, exhaled nitric oxide and clinical data from the Severe Asthma Research Program using unsupervised clustering, and then characterize them using supervised learning approaches. Methods Unsupervised clustering approaches were applied to 112 clinical, physiologic and inflammatory variables from 378 subjects. Variable selection and supervised learning techniques were employed to select relevant and nonredundant variables, address their predictive values, as well as the predictive value of the full variable set. Results Ten variable clusters and six subject clusters were identified, which differed and overlapped with previous clusters. Traditionally defined severe asthmatics distributed through subject Clusters 3–6. Cluster 4 identified early onset allergic asthmatics with low lung function and eosinophilic inflammation. Later onset, mostly severe asthmatics with nasal polyps and eosinophilia characterized Cluster 5. Cluster 6 asthmatics manifested persistent inflammation in blood and bronchoalveolar lavage and exacerbations despite high systemic corticosteroid use and side effects. Age of asthma onset, quality of life, symptoms, medications and health care utilization were some of the 51 nonredundant variables distinguishing subject clusters. These 51 variables classified test cases with 88% accuracy, compared to 93% accuracy with all 112 variables. Conclusion The unsupervised machine learning approaches used here provide unique insights into disease, confirming other approaches while revealing novel additional phenotypes.

Severe cases of coronavirus disease 2019 (COVID-19) cannot be adequately managed with mechanical ventilation alone. The role and outcome of extracorporeal membrane oxygenation (ECMO) in the management of COVID-19 is currently unclear. Eight COVID-19 patients have received ECMO support in Shanghai with seven with venovenous (VV) ECMO support and one veno arterial (VA) ECMO during cardiopulmonary resuscitation. As of March 25, 2020, four patients died (50% mortality), three patients (37.5%) were successfully weaned off ECMO after 22, 40, and 47 days support, respectively, but remain on mechanical ventilation. One patient is still on VV ECMO with mechanical ventilation. The partial pressure of oxygen/fractional of inspired oxygen ratio before ECMO initiation was between 54 and 76, and all were well below 100. The duration of mechanical ventilation before ECMO ranged from 4 to 21 days. Except the one emergent VA ECMO during cardiopulmonary resuscitation, other patients were on ECMO support for between 18 and 47 days. In conclusion, ensuring effective, timely, and safe ECMO support in COVID-19 is key to improving clinical outcomes. Extracorporeal membrane oxygenation support might be an integral part of the critical care provided for COVID-19 patients in centers with advanced ECMO expertise.

Background: 2019-Novel coronavirus (2019-nCoV) outbreaks create challenges for hospital laboratories because thousands of samples must be evaluated each day. Sample types, interpretation methods, and corresponding laboratory standards must be established. The possibility of other infections should be assessed to provide a basis for clinical classification, isolation, and treatment. Accordingly, in the present study, we evaluated the testing methods for 2019-nCoV and co-infections. Methods: We used a fluorescence-based quantitative PCR kit urgently distributed by the Chinese CDC to detect 8274 close contacts in the Wuhan region against two loci on the 2019-nCoV genome. We also analyzed 613 patients with fever who underwent multiple tests for 13 respiratory pathogens; 316 subjects were also tested for 2019-nCoV. Findings: Among the 8274 subjects, 2745 (33.2%) had 2019-nCoV infection; 5277 (63.8%) subjects showed negative results in the 2019-nCoV nucleic acid test (non-019-nCoV); and 252 cases (3.0%) because only one target was positive, the diagnosis was not definitive. Sixteen patients who originally had only one positive target were re-examined a few days later; 14 patients (87.5%) were finally defined as 2019-nCoV-positive, and 2 (12.5%) were finally defined as negative. The positive rates of nCoV-NP and nCovORF1ab were 34.7% and 34.7%, respectively. nCoV-NP-positive only and nCovORF1ab-positive cases accounted for 1.5% and 1.5%, respectively. In the 316 patients with multiple respiratory pathogens, 104 were positive for 2019-nCov and 6/104 had co-infection with coronavirus (3/104), influenza A virus (2/104), rhinovirus (2/104), and influenza A H3N2 (1/104); the remaining 212 patients had influenza A virus (11/202), influenza A H3N2 (11/202), rhinovirus (10/202), respiratory syncytial virus (7/202), influenza B virus (6/202), metapneumovirus (4/202), and coronavirus (2/202). Interpretation: Clinical testing methods for 2019-nCoV require improvement. Importantly, 5.8% of 2019-nCoV infected and 18.4% of non-2019-nCoV-infected patients had other pathogen infections. It is important to treat combined infections and perform rapid screening to avoid cross-contamination of patients. A test that quickly and simultaneously screens as many pathogens as possible is needed.