Brain-mimetic Kernel: A Kernel Constructed from Human fMRI Signals Enabling a Brain-mimetic Visual Recognition Algorithm

Hiroki Kurashige, Hiroyuki Hoshino, Takashi Owaki, Kenichi Ueno, Topi Tanskanen, Kang Cheng, Hideyuki Câteau

Research output: Chapter in Book/Report/Conference proceedingConference contributionScientificpeer-review

Abstract

Although the present-day machine learning algorithm sometimes beats humans in visual recognition, we still find significant differences between the brain’s and the machine’s visual processing. Thus, it is not guaranteed that the information sampled by the machines is the one used in the human brain. To overcome this situation, we propose a novel method for extracting the building blocks of our brain information processing and utilize them in machine learning algorithms. The visual features used in our brain are identifiable by applying kernel canonical correlation analysis (KCCA) to paired data of visual stimuli (images) and evoked functional magnetic resonance imaging (fMRI) activity. A machine learning algorithm incorporating the identified visual features represented in the brain is expected to inherit the characteristics of brain information processing. In the proposed method, the features are incorporated into kernel-based algorithms as a positive-definite kernel. Applying the method to fMRI data measured from a participant seeing natural and object images, we constructed a support vector machine (SVM) working on the visual features presumably used in the brain. We showed that our model outperforms the SVM equipped with a conventional kernel, especially when the size of the training data is small. Moreover, we found that the performance of our model was consistent with physiological observations in the brain, suggesting its neurophysiological validity.

Original languageEnglish
Title of host publicationNeural Information Processing : 28th International Conference, ICONIP 2021, Sanur, Bali, Indonesia, December 8–12, 2021, Proceedings, Part III
EditorsTeddy Mantoro, Minho Lee, Media Anugerah Ayu, Kok Wai Wong, Achmad Nizar Hidayanto
Number of pages13
Place of PublicationCham
PublisherSpringer Nature Switzerland AG
Publication date2021
Pages271-283
ISBN (Print)978-3-030-92237-5
ISBN (Electronic)978-3-030-92238-2
DOIs
Publication statusPublished - 2021
MoE publication typeA4 Article in conference proceedings
EventInternational Conference on Neural Information Processing - Virtual, Online
Duration: 8 Dec 202112 Dec 2021
Conference number: 28
https://iconip2021.apnns.org/

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume13110
ISSN (Print)0302-9743
ISSN (Electronic)1611-3349

Fields of Science

  • Brain-mimetic visual recognition algorithms
  • Functional magnetic resonance imaging
  • Kernel canonical correlation analysis
  • 3124 Neurology and psychiatry
  • 413 Veterinary science

Cite this