Anzeige
Erschienen in:
01.02.2024
Which design decisions in AI-enabled mobile applications contribute to greener AI?
Erschienen in: Empirical Software Engineering | Ausgabe 1/2024
EinloggenAktivieren Sie unsere intelligente Suche, um passende Fachinhalte oder Patente zu finden.
Wählen Sie Textabschnitte aus um mit Künstlicher Intelligenz passenden Patente zu finden. powered by
Markieren Sie Textabschnitte, um KI-gestützt weitere passende Inhalte zu finden. powered by
Abstract
Background
The usage of complex artificial intelligence (AI) models demands expensive computational resources. While currently, available high-performance computing environments can support such complexity, the deployment of AI models in mobile devices, which is an increasing trend, is challenging. Environments with low computational resources imply limitations in the design decisions during the AI-enabled software engineering lifecycle that balance the trade-off between the accuracy and the complexity of the mobile applications.
Objective
Our objective is to systematically assess the trade-off between accuracy and complexity when deploying complex AI models (e.g. neural networks) to mobile devices in pursuit of greener AI solutions. We aim to cover (i) the impact of the design decisions on the achievement of high-accuracy and low resource-consumption implementations; and (ii) the validation of profiling tools for systematically promoting greener AI.
Method
We implement neural networks in mobile applications to solve multiple image and text classification problems on a variety of benchmark datasets. We then profile and model the accuracy, storage weight, and time of CPU usage of the AI-enabled applications in operation with respect to their design decisions. Finally, we provide an open-source data repository following the EMSE open science practices and containing all the experimentation, analysis, and reports in our study.
Results
We find that the number of parameters in the AI models makes the time of CPU usage scale exponentially in convolutional neural networks and logarithmically in fully-connected layers. We also see the storage weight scales linearly with the number of parameters, while the accuracy does not. For this reason, we argue that a good practice for practitioners is to start small and only increase the size of the AI models when their accuracy is low. We also find that Residual Networks (ResNets) and Transformers have a higher baseline cost in time of CPU usage than simple convolutional and recurrent neural networks. Finally, we find that the dataset used for experimentation affects both the scaling properties and accuracy of the AI models, hence showing that researchers must study the presented set of design decisions in each specific problem context.
Conclusions
We have depicted an underlying and existing relationship between the design of AI models and the performance of the applications that integrate these, and we motivate further work and extensions to better characterize this complex relationship.