Single-input interpretation of a deep classification neural network

Scientific paper - Original scientific paper

• Author: Suraja Poštić
• Author: Marko Subašić
• Author's institution: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Electronic Systems and Information Processing)
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing; Artificial Intelligence

Abstract (english)

The justification for artificial intelligence-based predictions has become increasingly important due to their application in very delicate fields. In this work, we present MaxPIn as a linearization-based interpretation of deep classifiers that puts maximum emphasis on the input data values. The output conservation property of MaxPIn is achieved by including network and linearization bias contributions. Furthermore, we designed an interpretation technique specifically for the softmax layer,... More dubbed CCI, which is compatible with any output-conserving class score interpretation. The idea behind CCI is to consider input contributions to all the class scores simultaneously when calculating the final probability distribution interpretation for a single class. Our experiments show that MaxPIn achieves the most consistent quality regarding the assessment of input image area relevance compared to other popular interpretation methods. Including the network and linearization biases helped MaxPIn maintain the quality across different evaluation settings. We further show that zero-centered class score distributions with lower absolute values are more suitable for softmax extended interpretations. Finally, the comparison of interpretations at different network layers suggests that it is best to interpret the entire network. Less

• Keywords (english): ML interpretation; XAI; DCNNs; Image classification
• Language: english
• Publication type: Scientific paper - Original scientific paper
• Publication status: Published
• Peer review: Peer review - international
• Publication version: Accepted version
• Journal title: Knowledge-Based Systems
• Numbering: vol. 309, pp. 112903-112914
• BPC: 0 EUR
• p-ISSN: 0950-7051
• DOI: https://doi.org/10.1016/j.knosys.2024.112903
• URN:NBN: urn:nbn:hr:168:248448
• Publication: 2025-01-30
• Project: Number: KK.01.1.1.01.009; Title (croatian): DATACROSS - Napredne metode i tehnologije u znanosti o podatcima i kooperativnim sustavima; Title (english): Advanced methods and technologies in Data Science and Cooperative Systems; Acronym: DATACROSS; Leader: Sven Lončarić; Jurisdiction: eu; Funding stream: Cohesion Fund
• Document URL: https://linkinghub.elsevier.com/retrieve/pii/S0950705124015375
• Type of resource: Text
• Access conditions: Closed access
• Created on: 2025-03-25 11:15:25