Forward and Inverse Tasks for Inference Rate Estimation in Binary Neural Networks
Year: 2025
Type: article
Abstract: Deploying Binary Neural Networks (BNNs) presents significant challenges, particularly in selecting suitable hardware to achieve desired performance levels and accurately estimating computational costs. To address these issues, a novel metric named "XNOROP" was recently introduced, offering a simplified approach for estimating the computational cost of BNNs and introducing a method for compressing binary filters. This paper leverages " XNOROP" to define and solve two critical tasks in BNNs. The first task, referred to as the "Forward Task", involves estimating the inference rate of a given model M when deployed on a specific target device T. The second task, known as the "Inverse Task", outlines a systematic procedure to identify a set of target devices T capable of achieving a required inference rate M when deploying the model . Additionally, we extend the foundational formula of "XNOROP" and introduce “LXNOROP” which incorporates considerations for memory access time, enhancing its applicability for real-world deployment scenarios.
Authors: A. Shakkouf, V. S. Gromov
Institution: ITMO University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4413194741