Optimizing the Computational Offloading of Deep Neural Networks for Human Activity Recognition

Research Aim: Study the possibility of optimizing the computational offloading of deep neural networks by reducing the volume of data sent to the cloud with a focus on the application of human activity recognition with deep learning. Research method: In this research, three proposed methods of reducing the number of data samples, reducing the precision of data samples and compressing data samples are presented. In the first method, the data samples are deleted one in between or more before sending them. Data restoration in the cloud side is performed by interpolation estimates. In the precision reduction method, floating-point data samples are converted to integers with fewer precision before sending them. They are converted back on the cloud side by using the inverse conversion function. In the third method, the data is compressed with low overhead compression algorithms, either lossy or lossless, and is decompressed on the cloud side. Findings: Among the two proposed methods of reducing the number of samples and reducing the precision of data samples, both methods only slightly reduce the accuracy of activity detection. The latter method is superior to the former method due to a more significant reduction in data volume. Although the lossy compression method shows better results than the lossless method, neither is as effective as the precision reduction method and the reduction in the number of data samples. Conclusion: Practical results show that although the methods of reducing the number of samples and reducing their precision can decrease the volume of data sent without a significant effect on accuracy, the precision reduction method is superior due to greater data volume reduction. Furthermore, the delta compression method is not as effective as the other two methods.