Single-input interpretation of a deep classification neural network

Znanstveni rad - Izvorni znanstveni rad

• Autor: Suraja Poštić
• Autor: Marko Subašić
• Autorova ustanova: Sveučilište u Zagrebu; Fakultet elektrotehnike i računarstva; (Zavod za elektroničke sustave i obradbu informacija)
• Znanstveno / umjetničko područje, polje i grana: TEHNIČKE ZNANOSTI; Računarstvo; Umjetna inteligencija

Sažetak (engleski)

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,... Više 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. Sakrij dio sažetka

• Ključne riječi (engleski): ML interpretation; XAI; DCNNs; Image classification
• Jezik: engleski
• Vrsta publikacije: Znanstveni rad - Izvorni znanstveni rad
• Status objave: Objavljen
• Vrsta recenzije: Recenziran - međunarodna recenzija
• Verzija publikacije: Završna verzija rukopisa prihvaćena za objavljivanje (postprint)
• Naslov časopisa: Knowledge-Based Systems
• Brojčani podaci: vol. 309, str. 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
• Datum objave publikacije: 2025-01-30
• Projekt: Šifra: KK.01.1.1.01.009; Naziv (hrvatski): DATACROSS - Napredne metode i tehnologije u znanosti o podatcima i kooperativnim sustavima; Naziv (engleski): Advanced methods and technologies in Data Science and Cooperative Systems; Kratica: DATACROSS; Voditelj: Sven Lončarić; Pravna nadležnost: eu; Financijer: Europska unija; Linija financiranja: Cohesion Fund
• URL dokumenta: https://linkinghub.elsevier.com/retrieve/pii/S0950705124015375
• Vrsta resursa: Tekst
• Prava pristupa: Zatvoreni pristup
• Datum i vrijeme pohrane: 2025-03-25 11:15:25