RAVEN stands for a Reconfigurable Architecture for Varying Emerging Neural Networks. With the fast evolvement of neural network models, existing solutions lack the support for more complicated dataflow and operations, including reduction operation, sparse operation and nonlinear functions, etc. Properly leveraging the reconfigurability of the hardware, we are able to address those challenges while not introducing programming overhead, with a system diagram given below.
RAVEN was among the finalists of Qualcomm Innovation Fellowship in 2019. Currently, RAVEN is still in progress, and there are some related publications listed below.
- “UNO: Unify Nonlinear Operations via Approximate Taylor Series”, under review.
- “SECO: A Scalable Accuracy Approximate Exponential Function Via Cross-Layer Optimization”, in ISLPED 2019. [link]