On the benefits of self-taught learning for brain decoding

Area of Science:

• Neuroscience
• Machine Learning
• Data Science

Background:

• Functional magnetic resonance imaging (fMRI) generates complex statistical maps.
• Large public neuroimaging databases offer valuable resources for machine learning.
• Brain decoding aims to interpret neural activity related to cognitive processes.

Purpose of the Study:

• To investigate the benefits of self-taught learning using public fMRI data for brain decoding.
• To enhance the performance of classifiers on new tasks by leveraging pre-trained models.
• To assess the impact of data size and task complexity on model improvement.

Main Methods:

• Utilized the NeuroVault database of fMRI statistic maps.
• Trained a convolutional autoencoder for map reconstruction (pretraining).
• Initialized a supervised convolutional neural network with the pre-trained encoder for classification.

Main Results:

• Self-taught learning consistently improved classifier performance.
• Performance gains were dependent on the number of pretraining and fine-tuning samples.
• Model pretraining led to more generalizable features, reducing sensitivity to individual differences.

Conclusions:

• Pretraining on large neuroimaging datasets is beneficial for brain decoding.
• The effectiveness of this approach is influenced by data availability and task complexity.
• The method yields more robust and generalizable neural representations.