Latent hypernet: Exploring all Layers from Convolutional Neural Networks

  • 2018-11-16 14:57:28
  • Artur Jordao, Ricardo Kloss, William Robson Schwartz
  • 0

Abstract

Since Convolutional Neural Networks (ConvNets) are able to simultaneouslylearn features and classifiers to discriminate different categories ofactivities, recent works have employed ConvNets approaches to perform humanactivity recognition (HAR) based on wearable sensors, allowing the removal ofexpensive human work and expert knowledge. However, these approaches have theirpower of discrimination limited mainly by the large number of parameters thatcompose the network and the reduced number of samples available for training.Inspired by this, we propose an accurate and robust approach, referred to asLatent HyperNet (LHN). The LHN uses feature maps from early layers (hyper) andprojects them, individually, onto a low dimensionality space (latent). Then,these latent features are concatenated and presented to a classifier. Todemonstrate the robustness and accuracy of the LHN, we evaluate it using fourdifferent networks architectures in five publicly available HAR datasets basedon wearable sensors, which vary in the sampling rate and number of activities.Our experiments demonstrate that the proposed LHN is able to produce richinformation, improving the results regarding the original ConvNets.Furthermore, the method outperforms existing state-of-the-art methods.

 

Quick Read (beta)

loading the full paper ...