arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2008.02397cs.LGeess.IV
收藏

DANA:多变量传感器数据的自适应神经网络结构

DANA: Dimension-Adaptive Neural Architecture for Multivariate Sensor Data

Imperial College London(帝国理工学院);Queen Mary University of London(伦敦玛丽女王大学)

Mohammad Malekzadeh, Richard G. Clegg, Andrea Cavallaro, Hamed Haddadi

摘要:

可穿戴和移动设备中嵌入的运动传感器允许动态选择传感器流和采样率,从而实现多种应用,例如电源管理和数据共享控制。虽然深度神经网络(DNN)在传感器数据分类中实现了有竞争力的准确性,但DNN通常以固定的采样率处理来自固定传感器集合的传入数据,并且其输入维度的变化导致相当大的准确性损失、不必要的计算或操作故障。我们引入了一个维度自适应池(DAP)层,使DNN灵活,更强大的传感器可用性和采样率的变化。DAP对可变维度的卷积滤波器映射进行操作,并产生适合于前馈和递归层的固定维度的输入。我们还提出了一个维度自适应训练(DAT)过程,使使用DAP的DNN能够在推理时更好地概括可行的数据维度集。DAT包括在前向通道期间随机选择尺寸和利用若干后向通道的累积梯度进行优化。结合DAP和DAT,我们展示了如何将非自适应DNN转换为自适应神经架构(DANA),同时保持相同数量的参数。与现有的方法相比,我们的解决方案在推理时在可能的数据维度范围内提供了更好的分类精度,并且不需要上采样或插补,从而减少了不必要的计算。在七个数据集(四个用于人类活动识别的基准真实世界数据集和三个合成数据集)上的实验表明,DANA防止了最先进的DNN的分类准确性的显著损失,并且与基线相比,它在动态传感器可用性和不同的采样率下更好地捕获传感器数据中的相关模式。

英文摘要:

Motion sensors embedded in wearable and mobile devices allow for dynamic selection of sensor streams and sampling rates, enabling several applications, such as power management and data-sharing control. While deep neural networks (DNNs) achieve competitive accuracy in sensor data classification, DNNs generally process incoming data from a fixed set of sensors with a fixed sampling rate, and changes in the dimensions of their inputs cause considerable accuracy loss, unnecessary computations, or failure in operation. We introduce a dimension-adaptive pooling (DAP) layer that makes DNNs flexible and more robust to changes in sensor availability and in sampling rate. DAP operates on convolutional filter maps of variable dimensions and produces an input of fixed dimensions suitable for feedforward and recurrent layers. We also propose a dimension-adaptive training (DAT) procedure for enabling DNNs that use DAP to better generalize over the set of feasible data dimensions at inference time. DAT comprises the random selection of dimensions during the forward passes and optimization with accumulated gradients of several backward passes. Combining DAP and DAT, we show how to transform non-adaptive DNNs into a Dimension-Adaptive Neural Architecture (DANA), while keeping the same number of parameters. Compared to existing approaches, our solution provides better classification accuracy over the range of possible data dimensions at inference time and does not require up-sampling or imputation, thus reducing unnecessary computations. Experiments on seven datasets (four benchmark real-world datasets for human activity recognition and three synthetic datasets) show that DANA prevents significant losses in classification accuracy of the state-of-the-art DNNs and, compared to baselines, it better captures correlated patterns in sensor data under dynamic sensor availability and varying sampling rates.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑