Naslov
Showing the Impact of Data Augmentation on Model's Decisions Using Integrated Gradients

Autori
Hrga, Ingrid ; Ivašić-Kos, Marina

Vrsta, podvrsta i kategorija rada
Radovi u zbornicima skupova, cjeloviti rad (in extenso), znanstveni

Izvornik
Intelligent Sustainable Systems, Lecture Notes in Networks and Systems 578 / Nagar, A. K. - Singapur : Springer, 2023

Skup
6th World Conference on Smart Trends in Systems, Security and Sustainability (WorldS4 2022)

Mjesto i datum
London, Ujedinjeno Kraljevstvo, 24.08.2022. - 27.08.2022

Vrsta sudjelovanja
Predavanje

Vrsta recenzije
Međunarodna recenzija

Ključne riječi
XAI, Integrated Gradients, Data Augmentation

Sažetak
Deep neural network-based systems are increasingly performing tasks that can significantly impact various aspects of human life, such as medical diagnosis or financial risk assessment. Therefore, it is important to understand their decision-making process in order to determine whether those decisions may have been biased. Among the many factors that can influence a model’s decision is that of a data augmentation strategy. In this paper, we analyze how various augmentation techniques affect the image area that the model takes into account when classifying images. We use Integrated Gradients as a method to calculate the attribution of model’s output to its input features. Integrated Gradients results are clustered to determine the strategies the model uses to decide on the label, and to uncover possible changes in decision- making strategy due to the application of a particular augmentation. The results show that even when the accuracy is fairly uniform among the models, there may be a significant difference between the areas of the image to which they give more importance, as a consequence of the applied augmentation.

Izvorni jezik
Engleski

Znanstvena područja
Računarstvo, Informacijske i komunikacijske znanosti

POVEZANOST RADA