| 市原 有生希 | D, 中間発表 | 数理情報学（計算神経科学） | 池田 和司☆, | 田中 沙織, | 川鍋 一晃, | 杉本 徳和 |
| 榎原 学人 | D, 中間発表 | 数理情報学（計算神経科学） | 池田 和司☆, | 田中 沙織, | 川鍋 一晃, | 杉本 徳和 |
|
title: Decoding covert-speech-associated activity with 30-channel OPM-MEG system abstract: Speech brain–machine interfaces (BMIs) based on covert speech offer a promising approach to supporting communication by people with severe paralysis. However, high-performance speech BMIs still mainly rely on invasive recordings, motivating the development of safer and more broadly applicable noninvasive approaches. Optically pumped magnetometer magnetoencephalography (OPM-MEG) might provide a more flexible noninvasive MEG approach than conventional SQUID-MEG, but its feasibility for decoding covert-speech-associated activity remains unclear. In this study, we examined whether OPM-MEG could capture class-discriminative activity associated with covert speech. Brain magnetic fields were recorded by a 30-channel OPM-MEG system, comprising 15 dual-axis optically pumped magnetometers, as 18 healthy adults performed a task that included a no-speech Control condition and covert speech of two phrases. Electromyography (EMG) was recorded simultaneously to assess speech-related muscle activity. During covert speech, EMG waveforms showed no clear muscle activity, EMG-based decoding remained near chance level, and trial-wise EMG–OPM-MEG correlations were small in the low-gamma band (30–45 Hz; r = −0.021). In contrast, sensor-level low-gamma OPM-MEG features discriminated the Control condition from covert speech significantly above chance, even at the single-trial level (p < 0.01), and accuracy improved when test trials were averaged. Source-level analyses further suggested left-hemisphere-dominant decoding performance and above-chance decoding in ROIs broadly associated with motor and auditory/language processes. These findings provide initial evidence that even a relatively sparse 30-channel OPM-MEG configuration can capture covert-speech-associated class-discriminative activity that cannot be easily explained by speech-related peripheral muscle activity alone, demonstrating the potential of OPM-MEG as a noninvasive approach for future speech BMI research. language of the presentation: Japanese | ||||||
| 吉田 雄丸 | D, 中間発表 | 数理情報学（計算神経科学） | 池田 和司☆, | 田中 沙織, | 川鍋 一晃, | 杉本 徳和 |
|
title: Similarities in Top-Down Attention Between the Human Brain
abstract: Research comparing the human brain with deep neural networks (DNN) has revealed functional correspondences in visual and language processing. However, whether the two share a similar mechanism of top-down attention remains unexplored. This study investigates such similarities using Representational Similarity Analysis (RSA). Analysis was conducted using the fMRI dataset of Horikawa & Kamitani (2022), in which participants selectively attended to one of two superimposed objects. As a DNN, A pretrained ResNet50 was combined with a channel-attention module. To isolate attention-related representations, two refinements were introduced to standard RSA: (1) attention-based RSA, which restricts the RDM to image pairs sharing the same stimulus so that only attention-driven differences contribute. (2) a residual-vector formulation that treats the residual as an attentional signal. In addition to the above, three cue conditions (normal, reverse, random) were compared. In the attention-based RSA, the normal condition showed significantly higher correlation than the reverse and random conditions, indicating that DNN forms attention-modulated image representations resembling those of humans. In the residual-vector analysis, the normal and reverse conditions yielded significant correlations, suggesting that DNN encodes attentional representations partly independent of image features. These results provide initial evidence that DNN and the human brain share aspects of top-down attention. language of the presentation: Japanese | ||||||