Neural decoding of Aristotle tactile illusion using deep learning-based fMRI classification

• Deep learning-based fMRI classification was used to decode the Aristotle tactile illusion.¹
• Tactile decoding has indirect relevance to sensory neuroprosthetics.¹
• Study published in Frontiers.^{1 1}

Weekly enrichment (2026-07-20)

• The Frontiers in Neuroscience (2025) study applied three tactile stimuli (Aristotle, Reverse, Asynchronous) to the fingers of 30 participants during fMRI acquisition while recording how many stimuli each participant perceived.²
• Four convolutional neural network (CNN) models were trained for two task families: perception-based classification (based on number perceived) and stimulus-based classification (based on the physical stimulus applied).²
• The Simple Fully Convolution Network (SFCN) achieved the highest accuracy, reaching 68.4% for occurrence of the Aristotle vs. Reverse illusion and 80.1% for occurrence vs. absence of the Reverse illusion.^{2 3}
• Stimulus-based classification performed near chance (~50%), failing to distinguish among the three physical stimuli.^{2 3}
• Gradient-weighted class activation mapping (Grad-CAM) highlighted salient regions including the somatosensory cortex and parietal regions, plus the orbitofrontal cortex, middle temporal pole, supplementary motor area, and middle cingulate cortex.^{2 3}
• Severe class imbalance (544 illusory vs. 52 veridical trials) led the authors to exclude the occurrence-vs-absence of the Aristotle illusion from the main analysis to avoid majority-class bias.²
• The authors describe this as the first application of deep learning to fMRI acquired during tactile stimulation to decode the Aristotle illusion, concluding that perception-driven responses are more decodable than stimulus-driven ones.²
• Context: earlier somatosensory evoked potential work indicates early primary somatosensory cortex (S1) activity (~20 ms) reflects the perceived rather than the physical stimulus, while posterior parietal cortex activity (~200 ms) acts as a conflict resolver enabling veridical perception.⁴
• Context: the illusion is selectively reduced in focal hand dystonia, tying it to somatosensory cortical finger representation and supporting S1 involvement flagged by the CNN saliency maps.⁵

Footnotes

1. https://news.google.com/rss/articles/CBMilAFBVV95cUxNd1ZMNThET0c4TVJVMmJpSjhVeGZzY0FkUC11YVJnSi1JMzRHbTZySG5yTElxV3kzTzlDc0Zjcm5Eak1obnhqbEFMQ0IyR242b1dqV3VyMUlhdkdSU1NOcFlIeDVQWXdhdmd6OU00YkhyZDNrbE1xQkd2MVE2V1Y2YzVubm4wSVlyNHoyYmg5OFF5WU9o?oc=5 ↩ ↩² ↩³ ↩⁴

2. https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2025.1606801/full ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. https://pubmed.ncbi.nlm.nih.gov/40613086/ ↩ ↩² ↩³

4. https://doi.org/10.1093/cercor/bht037 ↩

5. https://doi.org/10.1093/brain/aws372 ↩