Give us a call or drop by anytime, we endeavor to answer all inquiries within 24 hours.
PO Box 16122 Collins Street West Victoria, Australia
firstname.lastname@example.org / email@example.com
Phone: + (066) 0760 0260 / + (057) 0760 0560
Deep learning (DL) is revolutionizing many fields. Major web companies are heavily relying on DL-based analytics, and there is excitement among the wider industries and the sciences for adopting DL. However, adopting DL in real-world applications is a non-trivial task. DL practitioners often have to iterate through complex stages of 1) data sourcing, 2) model building, and 3) model deploying. To overcome these bottlenecks, a new breed of specialized software systems, broadly referred to as Deep Learning Systems (e.g., TensorFlow, PyTorch, TVM), have emerged. The goal of these systems is to make DL adoption easier and efficient.
However, there is a significant limitation with the current generation DL systems: they don’t take into account the patterns and characteristics of end-to-end workloads. For example, exploratory nature is a common characteristic of many end-to-end DL workloads. In practice, this ends up generating workloads that launch several independent sub-tasks. However, these sub-tasks often overlap substantially on data and computations. A prime example is model selection, where one has to explore several different model configurations before picking the best one. As DL systems do not take into account such patterns, they often miss significant opportunities for optimization. On the contrary, relational database management systems, a much older subfield, has extensively explored how to optimize such workloads under the umbrella of multi-query optimization techniques.
To mitigate the above drawbacks of DL systems, we propose developing novel multi-query optimization-inspired techniques for optimizing end-to-end workloads in DL systems. In this presentation, I will present several examples of such optimization techniques that can accelerate and improve the resource efficiency of popular end-to-end DL workloads, spanning data sourcing, model building, and model deployment stages.
Supun Chathuranga Nakandala is inviting you to a scheduled Zoom meeting.
Topic: Supun Chathuranga Nakandala’s Personal Meeting Room
Join Zoom Meeting https://ucsd.zoom.us/j/7296707725?pwd=WXcwM0UvNFBoUkZIRVNJNVd6YnUydz09
Meeting ID: 729 670 7725
One tap mobile +16699006833,,7296707725# US (San Jose)
+12133388477,,7296707725# US (Los Angeles)
Dial by your location +1 669 900 6833 US (San Jose)
+1 213 338 8477 US (Los Angeles)
+1 669 219 2599 US (San Jose)
888 475 4499 US Toll-free
833 548 0276 US Toll-free
833 548 0282 US Toll-free
877 853 5257 US Toll-free
Meeting ID: 729 670 7725
Find your local number: https://ucsd.zoom.us/u/a1EzQv2RK