Recognizing a face, driving a car or a translation of a language are examples of complex forms of perception, understanding and control that are now able to be automated and performed by computational models, often with near human-levels of precision. But such recent breakthroughs from the field of deep learning, though, still exert demands on mobile and embedded devices that until recently was prohibitive – leaving algorithms of this type to the domain of cloud computing.
Algorithms typically exert severe demands on local device resources
As a consequence, familiar devices like fitness trackers, home sensors, smartwatches – and even smartphones until the last 18-months – have largely been left behind by the explosion of deep learning powered innovations because deep neural networks and their algorithms typically exert severe demands on local device resources that has prevented their adoption within mobile and embedded platforms.
Evolution of devices like phones and wearables
“Because sensor perception and reasoning are so fundamental to this class of computation, I believe the evolution of devices like phones, wearables and things will be crippled until we reach a point where current – and future – deep learning innovations can be simply and efficiently integrated into these systems”, says Nic Lane, Associate Professor at the University of Oxford visiting Helsinki and giving an open talk on machine learning systems and on-device AI at the University of Helsinki on Tuesday, February the 27th.
Developing general-purpose support for deep learning on resource-constrained mobile devices
Nic Lane specializes in the study of efficient machine learning under computational constraints, and over the last three years has pioneered a range of embedded and mobile forms of deep learning.
Critical enablers with transformative impact
In his coming lecture, he will describe the progress towards developing general-purpose support for deep learning on resource-constrained mobile and embedded devices that he sees as a critical enabler that will have a transformative impact similar to when affordable mobile network data first went mainstream.
“Primarily, this requires a radical reduction in the resources (viz. battery-life, memory and computation) consumed by these models – especially at inference time”, says Lane.
Nic Lane is highlighting various, largely complementary, approaches his team has invented to achieve this goal. These innovations include: binary “on-the-fly” networks, sparse layer representations, dynamic forms of compression, and scheduling partitioned model architectures.
Purpose-built deep learning accelerators emerging?
“Collectively, these techniques rethink how deep learning algorithms can execute not only to better cope with mobile and embedded device conditions; but also to increase the utilization of commodity processors (e.g., DSPs, GPUs, CPUs) – as well as emerging purpose-built deep learning accelerators”, he says.
Nic Lane’s talk “Machine Learning Systems: On-device AI and beyond” on Tuesday 27 February, 2018 at 3:15 pm at Small Hall, Main Building, University of Helsinki, Fabianinkatu 33, is the next one of the series “Helsinki Distinguished Lecture Series on Future Information Technology”. This year this series is hosted by Professor Giulio Jacucci from the Faculty of Science.
The event is free of charge and open to all interested in the leading research in information technology.
To attend, please register at https://elomake.helsinki.fi/lomakkeet/86777/lomake.html
Lane received the 2017 Google Faculty Award in machine learning and spent 4-years as a Lead Researcher at Microsoft Research based in Beijing. Read more on him here: http://niclane.org
Contact information: email@example.com
Minna Meriläinen-Tenhu, science communication, MinnaMeriTenhu, +358 50 415 0316, firstname.lastname@example.org