论文信息 - Towards Machine Learning with Zero Real-World Data

Towards Machine Learning with Zero Real-World Data

Machine Learning (ML) models are widely used to infer human activities. However, collecting data to train ML models in realworld often requires significant time and effort. In this paper, we suggest a novel data collection framework to leverage pre-built VR applications and simulating tools. In particular, we applied the concept of virtual Inertial Measurement Unit (IMU) to capture activities of an avatar in simulation. Our initial results show that Random Forest (RF), Support Vector Machine (SVM), and Long Short Term Memory (LSTM) models built with the virtual sensor data can classify three activities (i.e., standing, running, walking) over a realworld dataset at the accuracy of 80.40% (87.83% precision and 80.12% recall), 67.52% accuracy (72.24% precision and 68.15% recall), and 77.67% accuracy (86.25% precision and 77.63% recall), respectively. The early results show the initial feasibility of simulation-driven machine learning without real-world data

[1] D. K. Arvind,et al. IMUSim: A simulation environment for inertial sensing algorithm design and evaluation , 2011, Proceedings of the 10th ACM/IEEE International Conference on Information Processing in Sensor Networks.

[2] I. J. Schoenberg. Contributions to the problem of approximation of equidistant data by analytic functions. Part A. On the problem of smoothing or graduation. A first class of analytic approximation formulae , 1946 .

[3] Romain Tavenard,et al. Data Augmentation for Time Series Classification using Convolutional Neural Networks , 2016 .

[4] Varun Jampani,et al. Training Deep Networks with Synthetic Data: Bridging the Reality Gap by Domain Randomization , 2018, 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW).

[5] Razvan Pascanu,et al. Sim-to-Real Robot Learning from Pixels with Progressive Nets , 2016, CoRL.

[6] Sergey Levine,et al. Learning Neural Network Policies with Guided Policy Search under Unknown Dynamics , 2014, NIPS.

[7] Marcin Andrychowicz,et al. Sim-to-Real Transfer of Robotic Control with Dynamics Randomization , 2017, 2018 IEEE International Conference on Robotics and Automation (ICRA).

[8] Fernando Seco Granja,et al. Simulation of foot-mounted IMU signals for the evaluation of PDR algorithms , 2011, 2011 International Conference on Indoor Positioning and Indoor Navigation.