Error Distribution Modeling for Behavior-level Approximate Computing
TimeWednesday, July 13th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionOne challenge for Approximate Computing (AxC) is the lack of accurate error models. Existing error models can only describe static error statistics, which is largely inaccurate and uninformative. Therefore, we propose a novel error distribution model. We adopt Gaussian Mixture Model and propose a graph neural network (GNN) to predict the GMM parameters. Our approach has three contributions: (1) the distribution model is much more informative than error statistics; (2) it is more accurate in predicting error statistics; (3) our GNN-based predictor can generalize to unseen applications without retraining. Experiments demonstrate that our approach can accurately predict the error distribution.