|Title||:||Addressing the Computational Challenges of Ubiquitous Analytics and Learning|
|Speaker||:||Anand Raghunathan (IIT Madras & Purdue Univ.)|
|Details||:||Fri, 6 Jan, 2017 4:00 PM @ BSB 361|
|Abstract:||:||We are witnessing a profound change in the workloads that are driving the demand
for computing. In data centers and the cloud, computing demand is driven by the
need to organize, analyze, interpret, and search through exploding amounts of
data from the virtual and physical worlds. In mobile, wearable and IoT devices,
the need to make sense of and interact more intelligently with users and the
environment drive much of the computing demand. These trends have led to the
genesis of a new class of workloads for computing platforms that involve
recognition, mining, analytics, and inference. Machine learning is squarely at
the center of this trend indeed, the past decade has seen tremendous
developments in machine learning algorithms, and remarkable growth in their
In this talk, I will make a case for ubiquitous analytics and learning as a key driver in the design of future computing platforms. I will present a quantitative analysis of the computational requirements of deep learning networks a class of machine learning algorithms that have attracted great interest and achieved remarkable success in recent years. The analysis highlights a large gap between the capabilities of current computing systems and the requirements posed by these applications. This gap will only grow due to the seemingly insatiable appetite of these applications, together with diminishing benefits from technology scaling. I will outline a roadmap of technologies that can help bridge this gap accelerators for machine learning, approximate computing, neuromorphic hardware, and emerging post-CMOS devices.
This is a part of IITM-Purdue Seminar Series.