Control learning rate for autism facial detection via deep transfer learning.

Area of Science:

• Neuroscience
• Computer Science
• Medical Imaging

Background:

• Autism spectrum disorder (ASD) is a neurodevelopmental condition impacting social interaction and communication.
• Early ASD detection is crucial for improving individual outcomes.
• Machine learning, especially deep learning with convolutional neural networks (CNNs), shows promise for ASD diagnosis from facial images.

Purpose of the Study:

• To introduce a novel algorithm, the control subgradient algorithm (CSA), for optimizing deep CNNs in ASD diagnosis.
• To address the challenge of tedious learning rate selection in deep CNN training for ASD detection.
• To enhance the speed and accuracy of ASD classification using facial images.

Main Methods:

• Development of the control subgradient algorithm (CSA), a modified subgradient method with adaptive learning rate updates.
• Application of CSA to the DensNet-121 CNN model for facial ASD classification.
• Evaluation of CSA performance on a public facial ASD dataset, including comparisons with a baseline method and assessment with L2-regularization.

Main Results:

• The CSA algorithm demonstrated faster convergence compared to the baseline method.
• CSA improved classification accuracy and reduced loss in identifying ASD from facial images.
• The combination of CSA with L2-regularization further enhanced the performance of the deep CNN model.

Conclusions:

• The control subgradient algorithm (CSA) offers an effective and efficient approach for ASD diagnosis using deep CNNs and facial imagery.
• CSA provides a more optimized learning rate strategy, leading to improved diagnostic performance.
• This method holds potential for advancing automated ASD detection systems.