Autoencoder and restricted Boltzmann machine for transfer learning in functional magnetic resonance imaging task...

Area of Science:

• Neuroimaging
• Machine Learning
• Computational Neuroscience

Background:

• Deep neural networks (DNNs) are increasingly used for classifying functional magnetic resonance imaging (fMRI) data.
• Transfer learning is crucial for enhancing DNN performance and mitigating overfitting, especially with limited fMRI samples.

Purpose of the Study:

• To systematically compare autoencoder (AE) and restricted Boltzmann machine (RBM) for unsupervised pre-training of DNNs using resting-state fMRI (rfMRI) data.
• To evaluate two transfer learning schemes (weight fixing and fine-tuning) for DNN classifiers on task-based fMRI (tfMRI) data from the Human Connectome Project (HCP).

Main Methods:

• Unsupervised pre-training of DNNs using AE and RBM on rfMRI data.
• Implementation of two transfer learning strategies: fixing and fine-tuning pre-trained weights.
• Comparison of transfer learning DNNs against a baseline DNN trained with random initial weights on tfMRI classification tasks.

Main Results:

• Transfer learning generally outperformed the baseline DNN, with fixed RBM weights (14.8% error) and fine-tuned AE weights (15.1% error) showing superior performance.
• The optimal transfer learning strategy varied across different HCP task conditions.
• Fixed-weight transfer learning significantly reduced computational load compared to fine-tuning.

Conclusions:

• Weight initialization using RBM-based pre-trained weights is a promising approach for whole-brain fMRI task classification.
• The proposed AE/RBM-based pre-training scheme can improve classification performance and computational efficiency for various fMRI tasks.