Control Variate Approximation for DNN Accelerators

In this work, we introduce a control variate approximation technique for low error approximate Deep Neural Network (DNN) accelerators. The control variate technique is used in Monte Carlo methods to achieve variance reduction. Our approach significantly decreases the induced error due to approximate...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:2021 58th ACM/IEEE Design Automation Conference (DAC) S. 481 - 486
Hauptverfasser: Zervakis, Georgios, Spantidi, Ourania, Anagnostopoulos, Iraklis, Amrouch, Hussam, Henkel, Jorg
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 05.12.2021
Schlagworte:
Approximate Computing
Arithmetic Circuits
Control Variate
Convolution
Deep learning
Deep Neural Networks
Design automation
Low Power
MAC Array
Mathematical analysis
Monte Carlo methods
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In this work, we introduce a control variate approximation technique for low error approximate Deep Neural Network (DNN) accelerators. The control variate technique is used in Monte Carlo methods to achieve variance reduction. Our approach significantly decreases the induced error due to approximate multiplications in DNN inference, without requiring time-exhaustive retraining compared to state-of-the-art. Leveraging our control variate method, we use highly approximated multipliers to generate power-optimized DNN accelerators. Our experimental evaluation on six DNNs, for Cifar-10 and Cifar100 datasets, demonstrates that, compared to the accurate design, our control variate approximation achieves same performance and 24% power reduction for a merely 0.16% accuracy loss.
DOI:10.1109/DAC18074.2021.9586092