Publication
NPN-ApproxLib: DNN-Aware Approximate Multiplier Design via NPN Classes
Sallar Ahmadi-Pour; Chandan Jha; Rolf Drechsler
In: Proceedings of the 8th IEEE International Conference on Artificial Intelligence Circuits and Systems. IEEE International Conference on Artificial Intelligence Circuits and Systems (AICAS-2026), September 16-18, Ha Long Bay, Viet Nam, IEEE, 2026.
Abstract
Approximate circuits have shown immense potential
in error-resilient applications, with Deep Neural Networks (DNNs)
being one of the most prominent examples. Current works
explore approximate circuit errors only based on synthetic
input samples and isolated from their target application. While
automated approaches exist that explore a large number of
approximate designs by a specified bound on the circuit error,
it has recently been shown that this can lead to larger than
expected errors. Verifiable approximate circuits based on functional approximation alleviate this issue. However, approaches
that explore verifiable designs are limited due to the large design
space and therefore cannot evaluate the circuit in the target
application during the exploration. In this paper, we introduce
a methodology, based on Negation-Permutation-Negation (NPN)
classes, to explore a reduced number of approximate multipliers
based on the error within DNN applications to overcome limitations of previous works. Additionally, we will provide NPNApproxLib as an