Nowadays, deep learning methods with high structural complexity and flexibility inevitably lean on the computational capability of the hardware. A platform with high-performance GPUs and large amounts of memory could support neural networks having large numbers of layers and kernels. However, naively pursuing high-cost hardware would probably drag the technical development of deep learning methods. In the article, we thus establish a new preprocessing method to reduce the computational complexity of the neural networks. Inspired by the band theory of solids in physics, we map the image space into a noninteraction physical system isomorphically and then treat image voxels as particle-like clusters. Then, we reconstruct the Fermi–Dirac distribution to be a correction function for the normalization of the voxel intensity and as a filter of insignificant cluster components. The filtered clusters at the circumstance can delineate the morphological heterogeneity of the image voxels. We used the BraTS 2019 datasets and the dimensional fusion U-net for the algorithmic validation, and the proposed Fermi–Dirac correction function exhibited comparable performance to other employed preprocessing methods. By comparing to the conventional z-score normalization function and the Gamma correction function, the proposed algorithm can save at least 38% of computational time cost under a low-cost hardware architecture. Even though the correction function of global histogram equalization has the lowest computational time among the employed correction functions, the proposed Fermi–Dirac correction function exhibits better capabilities of image augmentation and segmentation.
Fusing object detection techniques and stochastic variational inference, we proposed a new scheme for lightweight neural network models, which could simultaneously reduce model sizes and raise the inference speed. This technique was then applied in fast human posture identification. The integer-arithmetic-only algorithm and the feature pyramid network were adopted to reduce the computational complexity in training and to capture features of small objects, respectively. Features of sequential human motion frames (i.e., the centroid coordinates of bounding boxes) were extracted by the self-attention mechanism. With the techniques of Bayesian neural network and stochastic variational inference, human postures could be promptly classified by fast resolving of the Gaussian mixture model for human posture classification. The model took instant centroid features as inputs and indicated possible human postures in the probabilistic maps. Our model had better overall performance than the baseline model ResNet in mean average precision (32.5 vs. 34.6), inference speed (27 vs. 48 milliseconds), and model size (46.2 vs. 227.8 MB). The model could also alert a suspected human falling event about 0.66 s in advance.
Fall accidents can cause severe impacts on the physical health and the quality of life of those who suffer limb diseases or injuries, the elderly, and their caregivers. Moreover, the later the accident is discovered, the lower the chance of recovery of the injured one. In order to detect accidents earlier, we propose a data-driven human fall detection framework. By combining the sensing mechanism of a commercialized webcam and an ultrasonic sensor array, we develop a probability model for automatic human fall monitoring. The webcam and ultrasonic array respectively collect the transverse and longitudinal time-series signals from a moving subject, and then these signals are assembled as a three-dimensional (3D) movement trajectory map. We also use two different detection-tracking algorithms for recognizing the tracked subjects. The mean height of the subjects is 164.2 ± 12 cm. Based on the data density functional theory (DDFT), we use the 3D motion data to estimate the cluster numbers and their cluster boundaries. We also employ the Gaussian mixture model as the DDFT kernel. Then, we utilize those features to build a probabilistic model of human falling. The model visually exhibits three possible states of human motions: normal motion, transition, and falling. The acceptable detection accuracy and the small model size reveals the feasibility of the proposed hybridized platform. The time from starting the alarm to an actual fall is on average about 0.7 s in our platform. The proposed sensing mechanisms offer 90% accuracy, 90% sensitivity, and 95% precision in the data validation. Then these vital results validate that the proposed framework has comparable performance to the contemporary methods.
Background: Containment measures slowed the spread of COVID-19 but led to a global economic crisis. We establish a reinforcement learning (RL) algorithm that balances disease control and economic activities. Methods: To train the RL agent, we design an RL environment with 4 semi-connected regions to represent the COVID-19 epidemic in Tokyo, Osaka, Okinawa, and Hokkaido, Japan. Every region is governed by a Susceptible-Exposed-Infected-Quarantined-Removed (SEIQR) model and has a transport hub to connect with other regions. The allocation of the synthetic population and inter-regional traveling is determined by population-weighted density. The agent learns the best policy from interacting with the RL environment, which involves obtaining daily observations, performing actions on individual movement and screening, and receiving feedback from the reward function. After training, we implement the agent into RL environments describing the actual epidemic waves of the four regions to observe the agent’s performance. Results: For all epidemic waves covered by our study, the trained agent reduces the peak number of infectious cases and shortens the epidemics (from 165 to 35 cases and 148 to 131 days for the 5th wave). The agent is generally strict on screening but easy on movement, except for Okinawa, where the agent is easy on both actions. Action timing analyses indicate that restriction on movement is elevated when the number of exposed or infectious cases remains high or infectious cases increase rapidly, and stringency on screening is eased when the number of exposed or infectious cases drops quickly or to a regional low. For Okinawa, action on screening is tightened when the number of exposed or infectious cases increases rapidly. Conclusions: Our experiments exhibit the potential of the RL in assisting policy-making and how the semi-connected SEIQR models establish an interactive environment for imitating cross-regional human flows.
19 March 2024 in IEEE Transactions on Emerging Topics in Computational Intelligence
Integrating frameworks of Fermi normalization and fast data density functional transform (fDDFT), we established a new global convolutional self-action module to reduce the computational complexity in modern deep convolutional neural networks (CNNs). The Fermi normalization conflates mathematical properties of sigmoid function and z-score normalization with high efficiency. Global convolutional kernels embedded in the fDDFT simultaneously extract global features from whole input images through long-range dependency. The fDDFT endows the transformed images with a smoothness property, so the images can be substantially down-sampled before the global convolutions and then resized back to the original dimensions without losing accuracy. To inspect the feasibility of the synergy of Fermi normalization and fDDFT and the combinational effect with modern CNNs, we applied the dimension-fusion U-Net as a backbone and utilized the datasets from BraTS 2020. Experimental results exhibited that the model embedded with the module saved 57%–60% computational costs and raised 50%–53% inferencing speeds compared to the naïve D-UNet model. Furthermore, the module enhanced the accuracy of brain tumor image segmentation. The dice scores of the work are 0.9221 for whole tumors, 0.8760 for tumor cores, 0.8659 for enhancing tumors, and 0.8362 for peritumoral edema. These results exhibit comparable performance to the winner of BraTS 2020. Our results also validate that image inputs processed by the module provide aligned and unified bases, establishing a specific space with optimized feature map combinations to reduce computational complexity efficiently. The module significantly boosted the performance of training and inferencing without losing model accuracy.