Deep Osmosis: Holistic Distributed Deep Learning in Osmotic Computing

Emerging availability (and varying complexity and types) of Internet of Things (IoT) devices, along with large data volumes that such devices (can potentially) generate, can have a significant impact on our lives, fuelling the development of critical next-generation services and applications in a variety of application domains (e.g. healthcare, smart grids, finance, disaster management, agriculture, transportation and water management). Deep learning technology, which has in the past been used successfully in computer vision and language modelling is now finding application in new domains driven by availability of diverse and large datasets. One such example is the advances in medical diagnostics and prediction by using Deep Learning technologies to improve human health. However, transferring large data streams (a requirement of Deep Learning technologies for achieving high accuracy) to centralised locations such as Cloud datacentre environments, in a timely and reliable manner, is being seen as a key limitation of expanding the application horizons of such technologies. To this end, various paradigms, including Osmotic Computing, have been proposed that promotes distribution of data analysis tasks across Cloud and Edge computing environments. However, these existing paradigms fail to provide a detailed account of how technologies such as deep learning can be orchestrated and take advantage of the cloud, edge and mobile edge environments in a holistic manner. In other words, the focus of this Blue Skies piece is to analyze the research challenges involved with developing a new class of holistic distributed deep learning algorithms that are “resource and data aware”, and which are able to account for underlying heterogeneous data and data models, resource (cloud vs. edge vs. mobile edge) models and data availability while executing – trading accuracy for execution time, etc.