Accelerating machine learning
Olivier Temam, responsable de l'action exploratoire ByMoore
Olivier Temam, Research Director at Inria, was awarded a Best Paper Award at the ASPLOS Conference, held in Salt Lake City from 1 to 5 March 2014. The reason for this award was his research into IT accelerators, which are infinitely more effective than traditional processors. This work has opened up new possibilities in the field of machine-learning.
The purpose of accelerators is to create more effective and more energy-efficient systems. These IT circuits, which are still restricted to certain categories of application, use much less energy than processors. They also provide much greater computing capacities at the same cost (or at a much lower cost for the same level of performance).
These accelerators have become Olivier Temam's favourite subject. His work, rewarded with a Best Paper Awardat the last ASPLOS conference, deals more specifically with the contribution of these accelerators to machine-learning. This relatively new scientific discipline explores the ways in which a system can perform tasks which are difficult or impossible to carry out using traditional algorithms. And the machines in question? Search engines, as well as image analysis and voice recognition software, robotics, etc. "We will always need processors for general tasks, but for tasks which require lengthy computing times, such as those carried out by Internet search engines, for example, we will increasingly integrate accelerators around the processors" , explained Olivier Temam.
The fact remains that, from an industrial point of view, it is inconceivable to create an accelerator for each type of programme. This led to the idea of limiting their development to algorithms which have an extensive scope of application. Within the framework of collaboration with the Institute of Computing Technology in Beijing, Olivier Temam became specifically interested in theDeep Neural Networksalgorithm. "It is currently used by a very wide variety of machine-learning applications", stated Olivier Temam."This makes it a perfect candidate for the creation of an accelerator."
Olivier Temam and his Chinese colleagues have shown that this accelerator, which is currently under development, could have similar performance levels to those of processors 100 times larger. They could therefore be used in everyday objects, such as smartphones and car navigation systems. The popularisation of machine learning has begun...
These articles could interest you:
To see more
- ASPLOS conference (Architectural Support for Programming Languages and Operating Systems)
- Best Paper Award for « A Small-Footprint High-Throughput Accelerator for Ubiquitous Machine-Learning »
- Olivier Temam's personal page
- Diannao team
Olivier Temam , directeur de recherche