IBM Uses Deep Learning to Train Raspberry Pi

　　Computations requiring high performance computing (HPC) power may soon be done in the palm of your hand thanks to work done this summer by IBM Research in Dublin, Ireland.

　　While scientists have come a long away in teaching machines how to process images for facial recognition and understand language to translate texts, IBM researchers focused on a different problem: how to use artificial intelligence (AI) techniques to forecast a physical process. In this case, the focus was on ocean waves, using traditional physics-based models driven by external forces, such as the rise and fall of tides, winds blowing in different directions, the depth and physical properties of water influence the speed and height of the waves.

　　HPC is normally essential to resolve the differential equations that encapsulate these physical processes and their relationships, and the expense often limits the spatial resolution, physical processes and time-scales that can be investigated by a real-time forecasting platform. In an interview with EE Times, IBM Research Senior Research Manager Sean McKenna said an HPC cluster using Big Iron has generally been the solution to dealing with the heavy computational load. IBM Research wanted to see if it could do the same work more quickly and more simply, he said.

　　The differential equations approach has developed over the course of a century or more, he said. Machine learning through AI is not rule based. “It's non-linear mapping of one input space to an output space," McKenna said. "That's what everything is in AI right now."

　　Researchers developed a deep-learning framework that provides a 12,000 percent acceleration over these physics-based models at comparable levels of accuracy. McKenna said the validated deep-learning framework can be used to perform real-time forecasts of wave conditions using available forecasted boundary wave conditions, ocean currents, and winds.

　　“The deep learning method is more of a black box," he said. "It's a little bit of paradigm shift."

　　Deep learning isn't about physical modeling and science to figure out what's leading to a set of results, it's about using engineering to solve a problem, and being able to do it more efficiently and faster, said McKenna. “We can build a model, train that model and put in on a more computationally-efficient device," he said.

　　What is clear are the significant benefits. Massively reducing the computational expense means simulations can be done on a Raspberry Pi rather HPC infrastructure.

　　The deep-learning framework was trained to forecast wave conditions at a case-study site at Monterey Bay, Calif., using the physics-based Simulating WAves Nearshore (SWAN) model to generate training data for the deep learning network. Driven by measured wave conditions, ocean currents from an operational forecasting system, and wind data, the model was run between the beginning of April 2013 and end of July 2017, generating forecasts at three-hour intervals to provide a total of 12,400 distinct model outputs. The study expands and builds on a collaboration between IBM Research-Ireland, Baylor University and the University of Notre Dame.

　　The deep learning model has yet to be deployed to a physical device, said McKenna, but the study demonstrates that the reduction in computational expense means the simulation of a physics model could be done an Raspberry Pi or any other low-end computing device that's trained by HPC.

　　“That opens up possibilities as to where that model can be deployed," McKenna said.

　　Being able to accurately forecast ocean wave heights and directions are a valuable resource for many marine-based industries as they often operate in harsh environments where power and computing facilities are limited. One scenario includes a shipping company using highly accurate forecasts to determine the best voyage route in rough seas to minimize fuel consumption or travel time. A surfer could get data localized to a specific beach to ride the best waves, said McKenna.

　　IBM Research's deep learning model could potentially be leveraged to use existing HPC infrastructure to train cheaper computing devices, even a smartphone, he said. “HPC resources are becoming more available in the cloud, so even if you don't own that resource you probably have access to it," he said.