A survey of uncertainty in deep neural networks Gawlikowski, Jakob; Tassi, Cedrique Rovile Njieutcheu; Ali, Mohsin ...
The Artificial intelligence review,
10/2023, Letnik:
56, Številka:
Suppl 1
Journal Article
Recenzirano
Odprti dostop
Over the last decade, neural networks have reached almost every field of science and become a crucial part of various real world applications. Due to the increasing spread, confidence in neural ...network predictions has become more and more important. However, basic neural networks do not deliver certainty estimates or suffer from over- or under-confidence, i.e. are badly calibrated. To overcome this, many researchers have been working on understanding and quantifying uncertainty in a neural network’s prediction. As a result, different types and sources of uncertainty have been identified and various approaches to measure and quantify uncertainty in neural networks have been proposed. This work gives a comprehensive overview of uncertainty estimation in neural networks, reviews recent advances in the field, highlights current challenges, and identifies potential research opportunities. It is intended to give anyone interested in uncertainty estimation in neural networks a broad overview and introduction, without presupposing prior knowledge in this field. For that, a comprehensive introduction to the most crucial sources of uncertainty is given and their separation into reducible model uncertainty and irreducible data uncertainty is presented. The modeling of these uncertainties based on deterministic neural networks, Bayesian neural networks (BNNs), ensemble of neural networks, and test-time data augmentation approaches is introduced and different branches of these fields as well as the latest developments are discussed. For a practical application, we discuss different measures of uncertainty, approaches for calibrating neural networks, and give an overview of existing baselines and available implementations. Different examples from the wide spectrum of challenges in the fields of medical image analysis, robotics, and earth observation give an idea of the needs and challenges regarding uncertainties in the practical applications of neural networks. Additionally, the practical limitations of uncertainty quantification methods in neural networks for mission- and safety-critical real world applications are discussed and an outlook on the next steps towards a broader usage of such methods is given.
Despite the importance of accurate assessment for low-density lipoprotein cholesterol (LDL-C), the Friedewald formula has primarily been used as a cost-effective method to estimate LDL-C when ...triglycerides are less than 400 mg/dL. In a recent study, an alternative to the formula was proposed to improve estimation of LDL-C. We evaluated the performance of the novel method versus the Friedewald formula using a sample of 5,642 Korean adults with LDL-C measured by an enzymatic homogeneous assay (LDL-CD). Friedewald LDL-C (LDL-CF) was estimated using a fixed factor of 5 for the ratio of triglycerides to very-low-density lipoprotein cholesterol (TG:VLDL-C ratio). However, the novel LDL-C (LDL-CN) estimates were calculated using the N-strata-specific median TG:VLDL-C ratios, LDL-C5 and LDL-C25 from respective ratios derived from our data set, and LDL-C180 from the 180-cell table reported by the original study. Compared with LDL-CF, each LDL-CN estimate exhibited a significantly higher overall concordance in the NCEP-ATP III guideline classification with LDL-CD (p< 0.001 for each comparison). Overall concordance was 78.2% for LDL-CF, 81.6% for LDL-C5, 82.3% for LDL-C25, and 82.0% for LDL-C180. Compared to LDL-C5, LDL-C25 significantly but slightly improved overall concordance (p = 0.008). LDL-C25 and LDL-C180 provided almost the same overall concordance; however, LDL-C180 achieved superior improvement in classifying LDL-C < 70 mg/dL compared to the other estimates. In subjects with triglycerides of 200 to 399 mg/dL, each LDL-CN estimate showed a significantly higher concordance than that of LDL-CF (p< 0.001 for each comparison). The novel method offers a significant improvement in LDL-C estimation when compared with the Friedewald formula. However, it requires further modification and validation considering the racial differences as well as the specific character of the applied measuring method.
In this study, we propose a semi-supervised learning method for spiking neural networks based on spike-timing-dependent plasticity (STDP). The spiking neural network structure of the proposed method ...incorporates teacher neurons and synapses, which serve the same purpose as real-life teachers, who ensure that the actions of their students do not transcend social norms. In the first stage of the proposed learning method, STDP-based supervised learning is applied. In the second stage, STDP-based unsupervised learning is conducted in the absence of any input signal to the teacher neuron. The proposed learning method classified handwritten characters with higher accuracy than the existing method. On the MNIST dataset, the proposed method was approximately 5%, 1%, and 3% more accurate than the conventional algorithm on 100, 400, and 1600 excitatory neurons, respectively.
Linezolid has antimycobacterial activity in vitro and is increasingly used for patients with highly drug-resistant tuberculosis.
We enrolled 41 patients who had sputum-culture-positive extensively ...drug-resistant (XDR) tuberculosis and who had not had a response to any available chemotherapeutic option during the previous 6 months. Patients were randomly assigned to linezolid therapy that started immediately or after 2 months, at a dose of 600 mg per day, without a change in their background regimen. The primary end point was the time to sputum-culture conversion on solid medium, with data censored 4 months after study entry. After confirmed sputum-smear conversion or 4 months (whichever came first), patients underwent a second randomization to continued linezolid therapy at a dose of 600 mg per day or 300 mg per day for at least an additional 18 months, with careful toxicity monitoring.
By 4 months, 15 of the 19 patients (79%) in the immediate-start group and 7 of the 20 (35%) in the delayed-start group had culture conversion (P=0.001). Most patients (34 of 39 87%) had a negative sputum culture within 6 months after linezolid had been added to their drug regimen. Of the 38 patients with exposure to linezolid, 31 (82%) had clinically significant adverse events that were possibly or probably related to linezolid, including 3 patients who discontinued therapy. Patients who received 300 mg per day after the second randomization had fewer adverse events than those who continued taking 600 mg per day. Thirteen patients completed therapy and have not had a relapse. Four cases of acquired resistance to linezolid have been observed.
Linezolid is effective at achieving culture conversion among patients with treatment-refractory XDR pulmonary tuberculosis, but patients must be monitored carefully for adverse events. (Funded by the National Institute of Allergy and Infectious Diseases and the Ministry of Health and Welfare, South Korea; ClinicalTrials.gov number, NCT00727844.).
High-resolution patterning of quantum dot (QD) films is one of the preconditions for the practical use of QD-based emissive display platforms. Recently, inkjet printing and transfer printing have ...been actively developed; however, high-resolution patterning is still limited owing to nozzle-clogging issues and coffee ring effects during the inkjet printing and kinetic parameters such as pickup and peeling speed during the transfer process. Consequently, employing direct optical lithography would be highly beneficial owing to its well-established process in the semiconductor industry; however, exposing the photoresist (PR) on top of the QD film deteriorates the QD film underneath. This is because a majority of the solvents for PR easily dissolve the pre-existing QD films. In this study, we present a conventional optical lithography process to obtain solvent resistance by reacting the QD film surface with diethylzinc (DEZ) precursors using atomic layer deposition. It was confirmed that, by reacting the QD surface with DEZ and coating PR directly on top of the QD film, a typical photolithography process can be performed to generate a red/green/blue pixel of 3000 ppi or more. QD electroluminescence devices were fabricated with all primary colors of QDs; moreover, compared to reference QD-LED devices, the patterned QD-LED devices exhibited enhanced brightness and efficiency.
The literature on network effects has popularized a hypothesis that competition between incompatible technologies results in the "winner-take-all" outcome. For the survival of the firm in this sort ...of competition, the installed base has been emphasized. We argue that the validity of this hypothesis depends on how customers interact with one another (e.g., if they exchange advice or files). In some interaction networks, customers influenced by their acquaintances may adopt a lagging technology even when a lead technology has built a large installed base. The presence of such a local bias facilitates the persistence of incompatibilities. When local bias cannot be sustained in other interaction networks, one technology corners the market. Our study suggests that overemphasizing the installed base, while ignoring network structure, could mislead practitioners.
This paper proposes a change detection algorithm on multi-spectral images based on feature-level U-Net. A low-complexity pan-sharpening method is proposed to employ not only panchromatic images, but ...also multi-spectral images for enhancing the performance of the deep neural network. The high-resolution multi-spectral (HRMS) images are then fed into the proposed feature-level U-Net. The proposed feature-level U-Net consists of two-stages: a feature-level subtracting network and U-Net. The feature-level subtracting network is used to extract dynamic difference images (DI) for the use of low-level and high-level features. By employing this network, the performance of change detection algorithms can be improved with a smaller number of layers for U-Net with a low computational complexity. Furthermore, the proposed algorithm detects small changes by taking benefits of both geometrical and spectral resolution enhancement and adopting an intensity-hue-saturation (IHS) pan-sharpening method. A modified of IHS pan-sharpening algorithm is introduced to solve spectral distortion problem by applying mean filtering in high frequency. We found that the proposed change detection on HRMS images gives a better performance compared to existing change detection algorithms by achieving an average F-1 score of 0.62, a percentage correct classification (PCC) of 98.78%, and a kappa of 61.60 for test datasets.
In the development of hybrid electric vehicles (HEVs), a series hybrid powertrain is mainly utilized in tracked vehicles to reduce energy consumption. In order to achieve high energy efficiency while ...maintaining the required driving performance, key design parameters of traction systems, such as transmission ratio and motor torque and power, need to be optimized. With the aim of effectively analyzing a complex track system, this article proposes an equivalent inertia model, which collectively represents the motion of each component of the tracked vehicle. The equivalent inertia model showed that the inertial effect was 34.8% higher than when the total mass of the vehicle was considered exclusively. Based on this inertia model, design objectives, such as energy efficiency and driving performance, were defined as quantified functions. Because of the balanced relationships between the objective functions, this study formulated a multiobjective optimization problem that includes motor stack length and transmission gear ratio as design variables. Based on the multiobjective optimization results, a Pareto front was obtained, which illustrates the balanced relationships between the objective functions. Comparing the initial HEV design, the optimum designs can improve energy efficiency and driving performance as a maximum of 13.0% and 2.9%, respectively.
This study proposes a bilateral attention U-Net with a dissimilarity attention gate (DAG) for change detection on remote sensing imageries. The proposed network is designed with a bilateral ...dissimilarity encoding for the DAG calculation to handle reversible input images, resulting in high detection rates regardless of the order of the two input images for change detection. The DAG exploits all the combinations of joint features to avoid spectral information loss fed into an attention gate on the decoder side. The effectiveness of the proposed method was evaluated on the KOMPSAT-3 satellite images dataset and the aerial change detection dataset (CDD). Its performance was better than that of conventional methods (specifically, U-Net, ATTUNet, and Modified-UNet++) as it achieved average F1-score and kappa coefficient (KC) values of 0.68 and 66.93, respectively, for the KOMPSAT-3 dataset. For CDD, it achieved F1-score and KC values of 0.70 and 68.74, respectively, which are also better values than those achieved by conventional methods. In addition, we found that the proposed bilateral attention U-Net can provide the same changed map regardless of whether the image order is reversed.
The aim of this study is to compare Friedewald-estimated and directly measured low density lipoprotein cholesterol (LDL-C) values and assess the concordance in guideline risk classification between ...the two methods.
The data were derived from the 2009 to 2011 Korea National Health and Nutrition Examination Survey. We included subjects with triglyceride (TG) levels < 400 mg/dL. Analysis was done for 6,454 subjects who had all lipid panels- total cholesterol, directly measured LDL-C, high density lipoprotein cholesterol (HDL-C), and TG.
The subjects ranged in age from 10 to 87 years old. The mean age was 41.5 ± 17.3 years. For subjects with TG < 400 mg/dL, overall concordance in guideline risk classification was 79.1%. The Friedewald formula tended to underestimate LDL-C more at higher TG or lower HDL-C levels. Especially, the percent of subjects who were misclassified into a lower risk category was 31% when TG were 200 to 299 mg/dL; and 45.6% when TG were 300 to 399 mg/dL. A greater underestimation of LDL-C occurred at higher TG and lower Friedewald-estimated LDL-C levels. Of subjects with a Friedewald-estimated LDL-C < 70 mg/dL, 55.4% had a directly measured LDL-C ≥ 70 mg/dL when TG were 200 to 399 mg/dL.
The Friedewald equation tends to underestimate LDL-C in highrisk subjects such as hypertriglyceridemia and hypo-HDL-cholesterolemia. For these individuals accurate assessment of LDL-C is crucial, and therefore additional evaluation is warranted.