Steve Homer | Automatically Scalable Computation

This is a talk organised by Ronald de Wolf that may be of interest to the ML group.
  • When May 19, 2016 from 04:00 PM to 05:00 PM (Europe/Amsterdam / UTC200)
  • Where L017
  • Add event to calendar iCal 
A system is proposed which speeds up some computations by taking advantage of current large, fast memory in order to efficiently store and recall relevant computational histories. The system automatically exploits predicted patterns in a computation in order to accelerate computation when it can.
Specifically our project is developing a system which,
- uses information compression methods to efficiently store full computation states in a large, fast cache,
- leverages and develops machine learning techniques to learn and analyze patterns in these computations, and finally
- speculatively collects and organizes resultant states and uses them to accelerate the original computation in many instances.

The theoretical side of this project considers models of computation to describe these physical systems and to better understand their capabilities. Our goal is to quantify the specific speed-ups that can be achieved and to provide lower bounds indicating the limits of our approach.

Joint work with Jonathan Appavoo, Thomas Unger (BU)
and David Brooks, Margo Seltzer, Amos Waterland (Harvard)