Salient Arithmetic Data Extraction from Brain Activity via an Improved Deep Network.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Computer Vision

Background:

• Automatic brain decoding requires interpreting neural activity from environmental stimuli.
• Analyzing visual stimulation brain recordings aids in understanding visual perception's effects on brain activity.

Purpose of the Study:

• To investigate the impact of arithmetic concepts on vision-related brain records.
• To propose an efficient Convolutional Neural Network-Generative Adversarial Network (CNN-GAN) for mapping electroencephalogram (EEG) to salient image features.

Main Methods:

• A CNN-GAN model was developed, with the CNN part utilizing depth-wise 1D convolutions for classifying brain signals.
• The CNN classifies EEG signals into 10 categories corresponding to Modified National Institute of Standards and Technology (MNIST) digits.
• The CNN output is fed into a fine-tuned GAN for image reconstruction.

Main Results:

• The CNN component achieved an average classification accuracy of 95.4% on 14-channel MindBigData EEG recordings.
• The CNN-GAN model demonstrated high performance with SSIM and CC saliency metrics of 92.9% and 97.28%, respectively.
• Successful EEG-based reconstruction of MNIST digits was achieved by fine-tuning the trained CNN-GAN weights.

Conclusions:

• The proposed CNN-GAN model effectively maps EEG signals to visual stimuli, enabling accurate classification and reconstruction of MNIST digits.
• This research advances brain decoding capabilities, particularly in understanding the relationship between visual perception and neural activity.
• The findings suggest significant potential for developing sophisticated brain-computer interfaces and applications in cognitive neuroscience.