Loading [MathJax]/extensions/MathZoom.js
Towards Green AI by Reducing Training Effort of Recurrent Neural Networks Using Hyper-Parameter Optimization with Dynamic Stopping Criteria | IEEE Conference Publication | IEEE Xplore

Towards Green AI by Reducing Training Effort of Recurrent Neural Networks Using Hyper-Parameter Optimization with Dynamic Stopping Criteria


Abstract:

Neural networks have become a leading model in modern machine learning, able to model even the most complex data. For them to be properly trained, however, a lot of compu...Show More

Abstract:

Neural networks have become a leading model in modern machine learning, able to model even the most complex data. For them to be properly trained, however, a lot of computational resources are required. With the carbon footprint of ever-growing adoption of neural networks in mind, an approach to reduce the required training resources would be very welcome. We designed a new training effort reduction method based on the calculation of area under the normalized loss curve and assessed it on the electricity consumption forecasting problem with the recurrent neural networks. The results show that the proposed method was able to considerably reduce the amount of computational resources, while maintaining the predictive performance, and thus contributing towards the Green AI.
Date of Conference: 19-21 September 2024
Date Added to IEEE Xplore: 05 November 2024
ISBN Information:

ISSN Information:

Conference Location: Pula, Croatia

I. Introduction

To meet the challenges of successfully discovering knowledge and intelligently analyzing the ever-increasing amount of data available, researchers have developed increasingly complex Machine Learning (ML) approaches, algorithms and models. Such approaches attempt to relieve data scientists of the increasingly demanding tasks of preprocessing and transforming data. A typical example are Neural Networks (NNs) and Deep Learning (DL) methods, which, through representation learning, largely relieve experts of the time-consuming feature engineering and mapping of data spaces [1]. The trade-off for this type of automation is an immense increase in the complexity of the training process and the computational complexity of such algorithms, which require huge amounts of computing resources to run successfully [2].

References

References is not available for this document.