A Probabilistic Re-Intepretation of Confidence Scores in Multi-Exit Models

In this paper, we propose a new approach to train a deep neural network with multiple intermediate auxiliary classifiers, branching from it. These ‘multi-exits’ models can be used to reduce the inference time by performing early exit on the intermediate branches, if the confidence of the prediction is higher than a threshold. They rely on the assumption that not all the samples require the same amount of processing to yield a good prediction. In this paper, we propose a way to train jointly all the branches of a multi-exit model without hyper-parameters, by weighting the predictions from each branch with a trained confidence score. Each confidence score is an approximation of the real one produced by the branch, and it is calculated and regularized while training the rest of the model. We evaluate our proposal on a set of image classification benchmarks, using different neural models and early-exit stopping criteria.     

The Existence and Global $p$-Exponential Stability of Periodic Solution for Stochastic BAM Neural Networks with Delays

In this paper, we consider a class of stochastic BAM neural networks with delays. By establishing new integral inequalities and using the properties of spectral radius of nonnegative matrix, some sufficient conditions for the existence and global $p$-exponential stability of periodic solution for stochastic BAM neural networks with delays are given. An example is provided to show the effectiveness of the theoretical results.     

Permutation Entropy-Based Interpretability of Convolutional Neural Network Models for Interictal EEG Discrimination of Subjects with Epileptic Seizures vs. Psychogenic Non-Epileptic Seizures

The differential diagnosis of epileptic seizures (ES) and psychogenic non-epileptic seizures (PNES) may be difficult, due to the lack of distinctive clinical features. The interictal electroencephalographic (EEG) signal may also be normal in patients with ES. Innovative diagnostic tools that exploit non-linear EEG analysis and deep learning (DL) could provide important support to physicians for clinical diagnosis. In this work, 18 patients with new-onset ES (12 males, 6 females) and 18 patients with video-recorded PNES (2 males, 16 females) with normal interictal EEG at visual inspection were enrolled. None of them was taking psychotropic drugs. A convolutional neural network (CNN) scheme using DL classification was designed to classify the two categories of subjects (ES vs. PNES). The proposed architecture performs an EEG time-frequency transformation and a classification step with a CNN. The CNN was able to classify the EEG recordings of subjects with ES vs. subjects with PNES with 94.4% accuracy. CNN provided high performance in the assigned binary classification when compared to standard learning algorithms (multi-layer perceptron, support vector machine, linear discriminant analysis and quadratic discriminant analysis). In order to interpret how the CNN achieved this performance, information theoretical analysis was carried out. Specifically, the permutation entropy (PE) of the feature maps was evaluated and compared in the two classes. The achieved results, although preliminary, encourage the use of these innovative techniques to support neurologists in early diagnoses.     

大脑血液动力学现象中的能量编码

神经信息的编码与解码是神经科学中的核心研究内容,同时又极具挑战性.传统的编码理论都具有各自的局限性,很难从脑的全局运行方式上给出有效的理论.而由于能量是一个标量具有可叠加性,因此能量编码理论可以从神经元活动的能量特征出发来研究脑功能的全局神经编码问题,取得了一系列的研究成果.本研究以王-张神经元能量计算模型为基础,构建了一个多层次结构的神经网络,通过计算机数值模拟得到了神经网络的能量消耗和血液中葡萄糖供能的变化情况.计算结果显示,和网络的神经活动达到峰值的时间相比,血液中葡萄糖的供能达到峰值的时间延迟了约5.6s.从定量的角度再现了功能性核磁共振(fMRI)中的血液动力学现象:大脑某个脑区的神经元集群被激活以后经过5~7 s的延迟,脑血流的变化才会大幅增加.模拟结果表明先前发表的由王-张神经元模型所揭示的负能量机制在控制大脑的血液动力学现象中起着核心的作用,预测了刺激条件下大脑的能量代谢与血流之间变化的本质是由神经元在发放动作电位过程中正、负能量之间的非平衡、不匹配性质所决定的.本文的研究结果为今后进一步探究血液动力学现象的生理学机制提供了新的研究方向,在神经网络的建模与计算方面给出了一个新的视角和研究方法.      

Deep Hierarchical Ensemble Model for Suicide Detection on Imbalanced Social Media Data

As a serious worldwide problem, suicide often causes huge and irreversible losses to families and society. Therefore, it is necessary to detect and help individuals with suicidal ideation in time. In recent years, the prosperous development of social media has provided new perspectives on suicide detection, but related research still faces some difficulties, such as data imbalance and expression implicitness. In this paper, we propose a Deep Hierarchical Ensemble model for Suicide Detection (DHE-SD) based on a hierarchical ensemble strategy, and construct a dataset based on Sina Weibo, which contains more than 550 thousand posts from 4521 users. To verify the effectiveness of the model, we also conduct experiments on a public Weibo dataset containing 7329 users’ posts. The proposed model achieves the best performance on both the constructed dataset and the public dataset. In addition, in order to make the model applicable to a wider population, we use the proposed sentence-level mask mechanism to delete user posts with strong suicidal ideation. Experiments show that the proposed model can still effectively identify social media users with suicidal ideation even when the performance of the baseline models decrease significantly.     

Automatic Search Dense Connection Module for Super-Resolution

The development of display technology has continuously increased the requirements for image resolution. However, the imaging systems of many cameras are limited by their physical conditions, and the image resolution is often restrictive. Recently, several models based on deep convolutional neural network (CNN) have gained significant performance for image super-resolution (SR), while extensive memory consumption and computation overhead hinder practical applications. For this purpose, we present a lightweight network that automatically searches dense connection (ASDCN) for image super-resolution (SR), which effectively reduces redundancy in dense connection and focuses on more valuable features. We employ neural architecture search (NAS) to model the searching of dense connections. Qualitative and quantitative experiments on five public datasets show that our derived model achieves superior performance over the state-of-the-art models.     

声纹的变压器放电与机械故障诊断研究

变压器中的一些放电和机械故障会产生异常声,可用于故障检测.据此,该文提出基于可听声的变压器放电和机械故障诊断方法.针对机械故障声与变压器本体噪声特征相似易混淆的问题,提出改进小波包-BP神经网络算法,与传统小波包-BP神经网络算法相比声音识别率提高了5.7％.为提高声音识别系统的泛化性,提出基于梅尔对数频谱和卷积神经网...     

Energy-Efficient Optimization for Energy-Harvesting-Enabled mmWave-UAV Heterogeneous Networks

Energy Harvesting (EH) is a promising paradigm for 5G heterogeneous communication. EH-enabled Device-to-Device (D2D) communication can assist devices in overcoming the disadvantage of limited battery capacity and improving the Energy Efficiency (EE) by performing EH from ambient wireless signals. Although numerous research works have been conducted on EH-based D2D communication scenarios, the feature of EH-based D2D communication underlying Air-to-Ground (A2G) millimeter-Wave (mmWave) networks has not been fully studied. In this paper, we considered a scenario where multiple Unmanned Aerial Vehicles (UAVs) are deployed to provide energy for D2D Users (DUs) and data transmission for Cellular Users (CUs). We aimed to improve the network EE of EH-enabled D2D communications while reducing the time complexity of beam alignment for mmWave-enabled D2D Users (DUs). We considered a scenario where multiple EH-enabled DUs and CUs coexist, sharing the full mmWave frequency band and adopting high-directive beams for transmitting. To improve the network EE, we propose a joint beamwidth selection, power control, and EH time ratio optimization algorithm for DUs based on alternating optimization. We iteratively optimized one of the three variables, fixing the other two. During each iteration, we first used a game-theoretic approach to adjust the beamwidths of DUs to achieve the sub-optimal EE. Then, the problem with regard to power optimization was solved by the Dinkelbach method and Successive Convex Approximation (SCA). Finally, we performed the optimization of the EH time ratio using linear fractional programming to further increase the EE. By performing extensive simulation experiments, we validated the convergence and effectiveness of our algorithm. The results showed that our proposed algorithm outperformed the fixed beamwidth and fixed power strategy and could closely approach the performance of exhaustive search, particle swarm optimization, and the genetic algorithm, but with a much reduced time complexity.