FTT-NAS: Discovering Fault-tolerant Convolutional Neural Architecture

With the fast evolvement of embedded deep-learning computing systems, applications powered by deep learning are moving from the cloud to the edge. When deploying neural networks (NNs) onto the devices under complex environments, there are various types of possible faults: soft errors caused by cosmic radiation and radioactive impurities, voltage instability, aging, temperature variations, malicious attackers, and so on. Thus, the safety risk of deploying NNs is now drawing much attention. In this article, after the analysis of the possible faults in various types of NN accelerators, we formalize and implement various fault models from the algorithmic perspective. We propose Fault-Tolerant Neural Architecture Search (FT-NAS) to automatically discover convolutional neural network (CNN) architectures that are reliable to various faults in nowadays devices. Then, we incorporate fault-tolerant training (FTT) in the search process to achieve better results, which is referred to as FTT-NAS. Experiments on CIFAR-10 show that the discovered architectures outperform other manually designed baseline architectures significantly, with comparable or fewer floating-point operations (FLOPs) and parameters. Specifically, with the same fault settings, F-FTT-Net discovered under the feature fault model achieves an accuracy of 86.2% (VS. 68.1% achieved by MobileNet-V2), and W-FTT-Net discovered under the weight fault model achieves an accuracy of 69.6% (VS. 60.8% achieved by ResNet-18). By inspecting the discovered architectures, we find that the operation primitives, the weight quantization range, the capacity of the model, and the connection pattern have influences on the fault resilience capability of NN models.

Multiple Connection Pattern Combination From Single-Mode Data for Mild Cognitive Impairment Identification

Mild cognitive impairment (MCI) is generally considered to be a key indicator for predicting the early progression of Alzheimer’s disease (AD). Currently, the brain connection (BC) estimated by fMRI data has been validated to be an effective diagnostic biomarker for MCI. Existing studies mainly focused on the single connection pattern for the neuro-disease diagnosis. Thus, such approaches are commonly insufficient to reveal the underlying changes between groups of MCI patients and normal controls (NCs), thereby limiting their performance. In this context, the information associated with multiple patterns (e.g., functional connectivity or effective connectivity) from single-mode data are considered for the MCI diagnosis. In this paper, we provide a novel multiple connection pattern combination (MCPC) approach to combine different patterns based on the kernel combination trick to identify MCI from NCs. In particular, sixty-three MCI cases and sixty-four NC cases from the ADNI dataset are conducted for the validation of the proposed MCPC method. The proposed method achieves 87.40% classification accuracy and significantly outperforms methods that use a single pattern.

Inter-Module Physical Interactions: A Force-Transmissive Modular Structure for Whole-Body Robot Motion

Robots are required to be significantly compliant and versatile to work in unstructured environments. In a number of studies, robots have positively exploited the environments during interactions and completed tasks from a morphological viewpoint. Modular robots can help realize real-world adaptive robots. Researchers have been investigating the actuation, coupling, and communication mechanisms among these robots to realize versatility. However, the diverse force transmission among modules needs to be further studied to achieve the adaptive whole-body dynamics of a robot. In this study, we fabricated a modular robot and proposed the realization of force transmission on this robot, by constructing fluid transferable network systems on the actuation modules. By exploiting the physical property variations of the modular robot, our experimental results prove that the robot’s motion can be changed by switching the connection pattern of the system.

Neural Correlates of Mental Workload During Multitasking: a Dynamic Causal Modeling Study

The goal of this study is to examine the neural correlates of different mental workload levels. Electroencephalogram (EEG) signals were recorded when participants perform a set of tasks simultaneously with low and high levels of mental workload. Brain connections for each workload level were estimated using Dynamic Causal Modeling (DCM), which is an effective connectivity method to reveal causal relationships between brain sources. The result showed a backward-only, left-lateralized connection pattern for high workload condition, compared to the bidirectional, two-sided connection pattern for low workload condition.These findings of the mental workload effect on neural mechanisms may be utilized in applications of the augmented cognition program.

SPATIAL CONNECTION CHARACTERISTICS AND DRIVING MECHANISM OF REGIONAL LOGISTICS IN GUANGDONG-HONG KONG-MACAO GREATER BAY AREA

The paper adopted the entropy weight TOPSIS method and the gravity model to explore the characteristics of logistics spatial connection of Guangdong-Hong Kong-Macao Greater Bay Area that includes nine prefecture-level cities and two special administrative regions, and analyzed the driving factors of the formation of logistics spatial connection pattern by geographical detector model. The results shows that: There are obvious imbalance in the comprehensive capacity of logistics in different regions, and the cities around the Pearl River estuary are generally strong in the logistics quality, like Guangzhou, Shenzhen and Hong Kong. The total amount of logistics links in different cities is significantly various. The logistics connection between cities are mainly weak, and the strong links are concentrated between Hong Kong and Shenzhen. In addition, Hong Kong, Guangzhou and Shenzhen are the three core nodes of the regional logistics network in Guangdong-Hong Kong-Macao Greater Bay Area. Industrial structure, economic scale, population scale, consumption level, the development of post and telecommunications industry are the main factors for the formation of the logistics spatial connection pattern. Moreover, these factors have a prominent driving effect on the cities with large amount of logistics links.

On twin preserving spanning subgraph

There are several approaches to lower the complexity of huge networks. One of the key notions is that of twin nodes, exhibiting the same connection pattern to the rest of the network. We extend this idea by defining a twin preserving spanning subgraph (TPS-subgraph) of a simple graph as a tool to compute certain graph related invariants which are preserved by the subgraph. We discuss how these subgraphs preserve some distance based parameters of the simple graph. We introduce a sub-skeleton graph on a vector space and examine its basic properties. The sub-skeleton graph is a TPS-subgraph of the non-zero component graph defined over a vector space. We prove that some parameters like the metric-dimension are preserved by the sub-skeleton graph.

Reconfigurable Patterns of Photovoltaic Cells to Reduce Power Loss due to Detrimental Shading Conditions

Poorly illuminated or defective photovoltaic cells (PV cells) affect the performance of the whole panel. In this work, we propose a methodology to discover the best connection pattern among cells in a panel under unfavorable shading conditions. It is, therefore, assumed that the target PV module allows internal disconnection and reconnection of neighboring cells. In order to manage the reconfiguration of the cells on the array, the developed algorithm searches for the connection pattern that yields the least power loss, taking into account the cell model in silicon and the effect of shading. Three progressive shading schemes have been applied to mimic possible hard shading on a 36-cell panel. For each case, the algorithm analyzed the IV-characteristics of the panel and suggested the best cell-connection pattern to achieve the highest power output under each condition. When less than 15 of the 36 PV cells were affected by shading, the algorithm was able to present up to 30% reduction in power loss when compared to the standard configuration, and up to 20% reduction when compared to complex fixed connection patterns. For shading patterns where 15 or more PV cells were affected, reconnection of the cells did not result in a reduction of power loss. However, if the PV panel is properly installed, only marginal hard shading should be expected and the algorithm would represent a promising tool to be deployed dynamically by means of switches and a management unit coupled to the panel.