CalibraSNN: Fair and Calibrated Convolutional Spiking Neural Network for High-Stakes Industry Applications

Authors:
Open Access
Publisher website
Google Scholar
Abstract
Real-world problems are often embedded in highly imbalanced contexts, where traditional machine learning algorithms struggle to achieve both strong performance and fairness. In high-stakes industries, incorporating sensitive attributes can enhance performance, but often at the cost of fairness. However, while artificial neural networks typically have the drawback of high energy usage, spiking neural networks (SNNs) offer a promising alternative because of their energy efficiency, in addition to their performance that is frequently hindered by their sensitivity to the hyperparameters. Our research introduces a fair and calibrated convolutional SNN modeling framework, CalibraSNN, designed for constrained, real-world problems by addressing the challenges in performance, fairness, energy usage, and sensitivity to hyperparameters. We evaluate our approach using the Bank Account Fraud dataset suite, which comprises real-world data and the business constraint of maintaining a false positive rate below 5%. Our results show up to 65% recall in fairly classifying fraudulent behaviors with competitive energy usage and power demand towards non-spiking approaches. The greener model is able to demand only 8 watts of power and use 30 joules with a recall of 54%. The experimental source code is available at https://github.com/dylanperdigao/CalibraSNN.
Funding Information
  • Portuguese Recovery and Resilience Plan (C645008882-00000055, Center for Responsible AI)

This publication has 35 references indexed in Scilit: