Spatio-temporal feature analysis of visual positional information on the mouse cortex using machine learning models
Type: dissertation
Abstract: Decoding visual information based on machine learning has the potential to reveal the neural mechanisms of visual information processing in the brain.Although most highly precise decoding methods require highly invasive electrodes, these electrodes make long-term use difficult because they cause brain damage.In this study, I examine the possibility of accurately decoding of visual stimuli from the neural activities recorded using low invasive electrode (ECoG).Recently, ECoG electrodes have attracted considerable attention (Alahi et al., 2021; Moon et al., 2024).Despite being low-invasive, they can record neural activity with high temporal resolution.However, the signals recorded by ECoG electrodes, known as local field potentials (LFPs), are composed of the collective neural activity of multiple neurons (Dubey & Ray, 2019;Buzsáki et al., 2012).LFPs obtained from adjacent recording channels tend to be reasonably redundant, resulting in low spatial resolution.Therefore, ECoG electrodes have mainly been adopted for humans and monkeys, which have wide cortical areas and small visual receptive fields.Contrary, the neurons in the mouse visual cortex have larger receptive field sizes and lower spatial resolution than those of primates (Table 1- 1, Schuett et al., 2002;Van den Bergh et al., 2010;Garrett et al., 2014;Zhuang et al., 2017;Kremkow et al., 2018).Furthermore, LFPs represent the integration of widespread cellular activity so that cellular functional characteristics might be averaged out.Therefore, it is a challenge to adopt ECoG electrodes for decoding visual stimuli in the mouse visual cortex.Our questions are as three follows: (1) How are the properties of neural signals recorded with ECoG electrodes in the mouse visual cortex?(2) Is it possible to decode visual information from the LFPs obtained by ECoG electrodes?(3) If it is possible, how is the visual information encoded in the spatio-temporal neural activity patterns?We presented a visual stimulus at multiple spatial positions to anesthetized mice and recorded evoked neural activities from the areas around the visual cortex using custom-made ECoG electrodes with a small number of recording sites.We then adopted the convolutional neural network (CNN) to train the association between the spatio-temporal patterns of the neural activity and the spatial positions of the visual stimuli.The decoding performance of the CNN model was examined by computing the discrimination accuracies with the cross-validation.We found that the CNN models enabled us to classify the stimulus positions with high accuracies even from the ECoG data of low spatial resolution and low dimension.Comparing the decoding performances to the CNN models trained by the surrogate data obtained by feature shuffling and dropping, we examined which features iii of the spatio-temporal neural activity patterns include relevant signals of the stimulus positions.The results suggest that the temporal variations of the LFPs are more informative than the spatial variations.
Author: 山田, 将太郎
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7223719647