Visual image reconstructed without semantics from human brain activity using linear image decoders and nonlinear...

Area of Science:

• Neuroscience
• Computer Vision
• Machine Learning

Background:

• Visual image reconstruction from brain activity has advanced, but current methods often rely on semantic information, potentially overlooking neural mechanisms.
• Existing techniques may not accurately represent true reconstruction from neural data due to a focus on semantic-to-image guidance.

Purpose of the Study:

• To develop a novel approach for reconstructing visual images from human brain activity that bypasses the need for semantic information.
• To enhance the accuracy and quality of visual reconstructions by addressing noise inherent in brain activity data.

Main Methods:

• A novel approach combining linear mapping with nonlinear noise suppression using a denoised deep convolutional neural network.
• Investigation included training shallow and deep autoencoder denoised neural networks, alongside a pre-trained state-of-the-art network.

Main Results:

• Combining linear image decoding with nonlinear noise reduction significantly improved the quality of reconstructed visual images from brain activity.
• The proposed methodology demonstrated effective decoding of perceptual experiences directly from brain activity patterns.

Conclusions:

• The developed method offers a promising pathway for decoding visual perception from neural data without semantic input.
• The model exhibits strong neural explanatory power due to its structural and functional similarities with the human visual brain.