Advancing neural network calibration: The role of gradient decay in large-margin Softmax optimization.

This study introduces a novel hyperparameter in the Softmax function to regulate the rate of gradient decay, which is dependent on sample probability. Our theoretical and empirical analyses reveal that both model generalization and calibration are significantly influenced by the gradient decay rate, particularly as confidence probability increases. Notably, the gradient decay varies in a convex or concave manner with rising sample probability. When employing a smaller gradient decay, we observe a curriculum learning sequence. This sequence highlights hard samples only after easy samples are adequately trained, and allows well-separated samples to receive a higher gradient, effectively reducing intra-class distances. However, this approach has a drawback: small gradient decay tends to exacerbate model overconfidence, shedding light on the calibration issues prevalent in modern neural networks. In contrast, a larger gradient decay addresses these issues effectively, surpassing even models that utilize post-calibration methods. Our findings provide substantial evidence that large margin Softmax can influence the local Lipschitz constraint by manipulating the probability-dependent gradient decay rate. This research contributes a fresh perspective and understanding of the interplay between large margin Softmax, curriculum learning, and model calibration through an exploration of gradient decay rates. Additionally, we propose a novel warm-up strategy that dynamically adjusts the gradient decay for a smoother L-constraint in early training, then mitigating overconfidence in the final model.

Radiology Residents' Perceptions of Artificial Intelligence: Nationwide Cross-Sectional Survey Study.

BACKGROUND: Artificial intelligence (AI) is transforming various fields, with health care, especially diagnostic specialties such as radiology, being a key but controversial battleground. However, there is limited research systematically examining the response of "human intelligence" to AI. OBJECTIVE: This study aims to comprehend radiologists' perceptions regarding AI, including their views on its potential to replace them, its usefulness, and their willingness to accept it. We examine the influence of various factors, encompassing demographic characteristics, working status, psychosocial aspects, personal experience, and contextual factors. METHODS: Between December 1, 2020, and April 30, 2021, a cross-sectional survey was completed by 3666 radiology residents in China. We used multivariable logistic regression models to examine factors and associations, reporting odds ratios (ORs) and 95% CIs. RESULTS: In summary, radiology residents generally hold a positive attitude toward AI, with 29.90% (1096/3666) agreeing that AI may reduce the demand for radiologists, 72.80% (2669/3666) believing AI improves disease diagnosis, and 78.18% (2866/3666) feeling that radiologists should embrace AI. Several associated factors, including age, gender, education, region, eye strain, working hours, time spent on medical images, resilience, burnout, AI experience, and perceptions of residency support and stress, significantly influence AI attitudes. For instance, burnout symptoms were associated with greater concerns about AI replacement (OR 1.89; P<.001), less favorable views on AI usefulness (OR 0.77; P=.005), and reduced willingness to use AI (OR 0.71; P<.001). Moreover, after adjusting for all other factors, perceived AI replacement (OR 0.81; P<.001) and AI usefulness (OR 5.97; P<.001) were shown to significantly impact the intention to use AI. CONCLUSIONS: This study profiles radiology residents who are accepting of AI. Our comprehensive findings provide insights for a multidimensional approach to help physicians adapt to AI. Targeted policies, such as digital health care initiatives and medical education, can be developed accordingly.

Posture adjustment and robust microinjection of zebrafish larval heart.

Unlike cells or embryos, zebrafish have a complex physiological structure, which poses challenges to posture recognition and adjustment during microinjection. Furthermore, zebrafish surface pigments exhibit strong interference with visual servo-based injection control, thus, affecting the success of microinjection and the subsequent survival rate. To address these challenges, we developed an automated microinjection system for the zebrafish heart that has advantages of high accuracy and success rate and avoids biological sample contamination. A convolutional neural networks (CNN) deep learning model is employed to determine the body axis posture. To solve the problems of blocked needle and abnormal tip positioning induced by zebrafish surface pigment during the injection process, an adaptive robust Kalman filter is proposed to suppress the abnormal values of visual feedback. Experimental results show that the success rate of body axis recognition based on the employed deep learning model exceeds 95%, and the proposed adaptive Kalman filter effectively suppresses the visual outliers, satisfying the requirements of high-precision injection for the zebrafish heart.

Association between extreme heat and hospital admissions for cataract patients in Hefei, China.

Cataract is the first cause of blindness and the major cause of visual impairment worldwide. Under conditions of global warming, researchers have begun to give attention to the influence of increasing temperature on cataract patients. Our paper aimed to investigate the association between extreme heat and hospital admissions for cataract in Hefei, China. Based on data from the New Rural Cooperative Medical System and National Meteorological Information Center, we used a generalized additive model and a distributed lag nonlinear model to examine the relationship between extreme heat and hospitalizations for cataract, with consideration of cumulative and lagged effects. When current mean temperature was above 28 °C, each 1 °C rise was associated with a 4% decrease in the number of cataract admissions (RR = 0.96, 95% CI = 0.94-0.98). The cumulative relative risk over 11 days of lag was the lowest, which indicated that every 1 °C increase in mean temperature above 28 °C was associated with a 19% decrease in the number of hospital admissions for cataract (RR = 0.81, 95% CI = 0.75-0.88). In subgroup analyses, the negative association between extreme heat and hospital admissions for cataract was stronger among patients who were not admitted to provincial-level hospitals. In conclusion, this paper found that extreme heat was negatively associated with cataract hospitalizations in Hefei, providing useful information for hospitals and policymakers.

Distributed Robust Filtering for Wireless Sensor Networks with Markov Switching Topologies and Deception Attacks.

This paper is concerned with the distributed full- and reduced-order l 2 - l ∞ state estimation issue for a class of discrete time-invariant systems subjected to both randomly occurring switching topologies and deception attacks over wireless sensor networks. Firstly, a switching topology model is proposed which uses homogeneous Markov chain to reflect the change of filtering networks communication modes. Then, the sector-bound deception attacks among the communication channels are taken into consideration, which could better characterize the filtering network communication security. Additionally, a random variable obeying the Bernoulli distribution is used to describe the phenomenon of the randomly occurring deception attacks. Furthermore, through an adjustable parameter E, we can obtain full- and reduced-order l 2 - l ∞ state estimator over sensor networks, respectively. Sufficient conditions are established for the solvability of the addressed switching topology-dependent distributed filtering design in terms of certain convex optimization problem. The purpose of solving the problem is to design a distributed full- and reduced-order filter such that, in the presence of deception attacks, stochastic external interference and switching topologies, the resulting filtering dynamic system is exponentially mean-square stable with prescribed l 2 - l ∞ performance index. Finally, a simulation example is provided to show the effectiveness and flexibility of the designed approach.

Temperature field model and control strategy in gravity casting process.

Temperature control is one of the most important processes during aluminum (Al) alloy engine cylinder head product casting. An improper temperature control may result in no uniformity and microstructure defects in casting parts and give rise to high defect ratio. In this paper, a mathematical model with high nonlinearity, strong coupling, and less uncertainty is developed for the solidification process in Al alloy casting. The interfacial heat transfer coefficient is combined with the mold structure comprehensively to build the temperature-structure model, and the characteristics of the uncertainty conversion are also used in order to achieve optimal temperature control during the solidification process. The cloud model integrated with Proportion-Integral-Differential (PID) temperature control system enables evaluation of the uncertainty conversion quantitatively. By inputting the temperature error and the temperature error rate, the PID inference is output through the cloud inference engine to achieve the optimal temperature curve. The superiority of the control algorithm was verified on a customized experimental platform with the temperature control system. Compared with manual operation and traditional PID control, the result shows that the error of the cloud model control is lower than the manual operation and traditional PID control. The experimental results also suggest that the performance of our cloud model is better than that of the manual operation model and the traditional PID control model regarding to stability and controllability.

A new optimal sliding mode controller design using scalar sign function.

This paper presents a new optimal sliding mode controller using the scalar sign function method. A smooth, continuous-time scalar sign function is used to replace the discontinuous switching function in the design of a sliding mode controller. The proposed sliding mode controller is designed using an optimal Linear Quadratic Regulator (LQR) approach. The sliding surface of the system is designed using stable eigenvectors and the scalar sign function. Controller simulations are compared with another existing optimal sliding mode controller. To test the effectiveness of the proposed controller, the controller is implemented on an aluminum beam with piezoceramic sensor and actuator for vibration control. This paper includes the control design and stability analysis of the new optimal sliding mode controller, followed by simulation and experimental results. The simulation and experimental results show that the proposed approach is very effective.