You are here
ECE Seminar: Succinct and Assured Deep Learning: Training and Execution
Tuesday, November 13, 2018
12:00 pm - 1:00 pm
Fitzpatrick Center Schiciano Auditorium Side B, room 1466
Bita Rouhani, a research scientist at Microsoft Research
One of the key challenges facing wide-scale adoption of Deep Learning (DL) is making the existing models more scalable, energy efficient, and reliable. This challenge is especially apparent on embedded edge devices where memory storage, battery life, and communication bandwidth are limited. A recent popular line of research has focused on performance optimization and DL acceleration using hardware and software co-design in an attempt to address this problem. My research work advances the state-of-the-art in this growing ¿eld by advocating a holistic co-design approach which not only includes hardware and software but also considers the geometry of the data, the underlying learning model, as well as safety concerns (e.g., robustness against adversarial attacks). I introduced, developed, and automated a resource-aware DL framework that achieves orders-of-magnitude energy ef¿ciency in the training and execution of deep learning models by simultaneous co-optimization of DL graph traversal, data embedding, and resource allocation. The proposed solution uses hardware profiling to identify the performance bottleneck(s) on the target platform and introduces extensible data and model transformation techniques that are aware of the pertinent resource provisioning and can adapt themselves accordingly. Such a co-optimization, in turn, enabled the first demonstration of succinct DL training on edge devices without the overhead of communicating with the cloud.