5.5.22-Efficient-AI-Michael-Carbin

• Video details

  The cost of training modern deep learning systems, such as GPT-3, has put the impressive capabilities of these systems beyond the reach of many individuals and institutions.  However, a key property of these systems is that they are approximate in that there is a natural trade-off between the quality of the results these systems produce and their performance and energy consumption. Exploiting this fact, researchers have developed a variety of new mechanisms that automatically change the exact behavior of a system to enable the system to execute more efficiently and cost effectively through techniques like quantization, distillation, and pruning. In this talk, I will present how such approximation mechanisms serve as a central opportunity for efficient ML, with still critical work to be done in understanding how to compose, analyze, and characterize the behavior of the resulting approximate systems.

Locked Interactive transcript