A semi-supervised classification RBM with an improved fMRI representation algorithm

Area of Science:

• Neuroimaging
• Machine Learning
• Computer Science

Background:

• Acquiring and labeling functional magnetic resonance imaging (fMRI) data for supervised learning is challenging.
• Semi-supervised learning leverages unlabeled data to improve feature learning and classification model performance.

Purpose of the Study:

• To develop an effective and robust semi-supervised learning classifier for fMRI data.
• To address limitations posed by insufficient labeled neuroimaging samples.

Main Methods:

• Proposed a hybrid L1/L2 regularization method (HRBM) for improved Restricted Boltzmann Machine (RBM) based fMRI representation.
• Developed a novel semi-supervised classification RBM (Semi-HRBM) using a joint training algorithm with HRBM.
• Integrated feature learning and classification into a single, optimized training process.

Main Results:

• The HRBM demonstrated satisfactory feature representation capabilities for fMRI data.
• The Semi-HRBM model improved average accuracy by 7.68% and average F1 score by 8.90% in a four-visual-stimuli classification task.
• Enhanced model generalization ability for fMRI data classification.

Conclusions:

• The Semi-HRBM model offers a valuable solution for studies with limited labeled neuroimaging data.
• This approach can aid in identifying complex brain states associated with various stimuli or tasks.