The research range for this fellowship is broad and based on Qualcomm’s core business and research areas. The winners are judged through a rigorous process that includes submission of a research abstract, a research proposal, and a presentation and a poster session. This year, the fellowship has been awarded to 13 teams from a total of 115 submissions from 22 schools. 

Crafton's and Chang’s proposal is entitled “Enabling Efficient Training of Deep Neural Networks through Sparse Direct Feedback Alignment and Algorithm-Hardware Co-Design.” This proposes a bio-plausible alternative to back-propagation of errors for real-time learning in neural networks. Here the error computation at a single synapse is local and avoids the weight propagation problem. 

Based partly on Crafton’s prior work on sparse feedback alignment and Chang’s work on near-memory computing architectures and IC design, the team demonstrated the possibility of reducing data-movement by orders of magnitude for classification and regression problems. This enables local learning rules based on a single error propagation, fast and parallel updates of all synapses in the network, and high energy-efficiency when realized in hardware.