eduzhai > Applied Sciences > Engineering >

DANA Dimension-Adaptive Neural Architecture for Multivariate Sensor Data

  • king
  • (0) Download
  • 20210507
  • Save

... pages left unread,continue reading

Document pages: 14 pages

Abstract: Motion sensors embedded in wearable and mobile devices allow for dynamicselection of sensor streams and sampling rates, enabling useful applications,e.g. for power management or control of data sharing. While deep neuralnetworks (DNNs) achieve competitive accuracy in sensor data classification,current DNN architectures only process data coming from a fixed set of sensorswith a fixed sampling rate, and changes in the dimensions of their inputs causeconsiderable accuracy loss, unnecessary computations, or failure in operation.To address this problem, we introduce a dimension-adaptive pooling (DAP) layerthat makes DNNs robust to temporal changes in sampling rate and in sensoravailability. DAP operates on convolutional filter maps of variable dimensionsand produces an input of fixed dimensions suitable for feedforward andrecurrent layers. Building on this architectural improvement, we propose adimension-adaptive training (DAT) procedure to generalize over the entire spaceof feasible data dimensions at the inference time. DAT comprises the randomselection of dimensions during the forward passes and optimization withaccumulated gradients of several backward passes. We then combine DAP and DATto transform existing non-adaptive DNNs into a Dimension-Adaptive NeuralArchitecture (DANA) without altering other architectural aspects. Our solutiondoes not need up-sampling or imputation, thus reduces unnecessary computationsat inference time. Experimental results, on four benchmark datasets of humanactivity recognition, show that DANA prevents losses in classification accuracyof the state-of-the-art DNNs, under dynamic sensor availability and varyingsampling rates.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×