METAHACI: Meta-learning for Human Activity Classification from IMU Data

Abstract

In the digital era, time-series data is pervasive to various applications and domains such as robotics, healthcare, finance, sport, etc. The inertial measurement unit (IMU) sensor is one of the popular devices collecting time-series data. Together with deep neural network implementation, this results in facilitating advancement in time series data analysis. However, the classical problem for the deep neural network is that it requires a vast amount of data which causes difficulty in the development and analysis process. Thus, we hypothesize that this problem can be avoided by combining data from many subjects that perform the same common tasks to get a larger amount of data. The consequence problem is that this decreases the overall classification accuracy. To tackle this, optimization-based meta-learning algorithms were selected, which are Reptile and model-agnostic meta-learning (MAML). The differences between Reptile and MAML are they use different rules for update gradient descent and differentiating through the optimization process. Leveraging their pre-existing meta-parameter weight before fine-tuning results in preferable accuracy for the existing amount of data, the figure is 89.65% for MAML and 78.38% for Reptile. To our knowledge, this is the first time comparing joint training with the meta-learning model, specifically, Reptile, and MAML for human activity classification from IMU sensors data. Experiment results show the superiority of classification accuracy from METAHACI proposed compare to joint training and simple CNNs.