Decoding natural images from evoked brain activities using encoding models with invertible mapping.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Encoding models map visual stimuli features to brain activity in the early visual cortex.
• Current models lack reversibility, preventing direct decoding of visual features from brain activity.

Purpose of the Study:

• To develop an invertible encoding model for the early visual cortex.
• To enable direct decoding of visual features from brain activity.
• To improve brain activity identification accuracy using fMRI data.

Main Methods:

• Designed a sparse framework-based encoding model predicting brain activity from complete feature representation.
• Introduced three key transformations based on primary visual cortex (V1) neuron rules to enhance features.
• Developed an invertible mapping using a closed-form formula.
• Implemented a hybrid identification method using Support Vector Machine (SVM) on fMRI data.

Main Results:

• The encoding model's mapping was successfully inverted.
• Identification accuracies for two subjects increased significantly from 92% and 72% to 98% and 92%.
• The chance level for identification was as low as 0.8%.

Conclusions:

• The developed encoding model is rational and effective for visual cortex studies.
• The invertible mapping facilitates direct decoding of visual features.
• The hybrid SVM method shows high accuracy in fMRI-based identification.