Hyper-Parameter Optimization of Stacked Asymmetric Auto-Encoders for Automatic Personality Traits Perception

Area of Science:

• Artificial Intelligence
• Machine Learning
• Speech Processing

Background:

• Deep learning models demonstrate potential for personality perception from speech.
• Manual hyper-parameter tuning is time-consuming, requires expertise, and can lead to suboptimal results.
• Existing optimization methods struggle with the vast hyper-parameter search space, risking local optima.

Purpose of the Study:

• To propose an automatic hyper-parameter tuning method for stacked asymmetric auto-encoders.
• To address the limitations of manual tuning and existing optimization techniques in deep learning.
• To enhance the accuracy and efficiency of personality perception from speech using deep learning.

Main Methods:

• Developed a novel global optimization method combining cultural algorithm, multi-island strategy, and parallelism.
• Evaluated the optimization method on benchmark datasets, comparing convergence speed and precision.
• Applied the method to tune five hyper-parameters of an asymmetric auto-encoder for personality perception, incorporating a novel cost function to prevent overfitting/underfitting.

Main Results:

• The proposed optimization method demonstrated faster convergence and improved precision on benchmark tests, effectively escaping local optima.
• Optimizing the asymmetric auto-encoder for personality perception resulted in a 6.52% improvement in unweighted average recall and a 9.54% improvement in accuracy.
• The method achieved remarkable outcomes compared to previous personality perception studies.

Conclusions:

• The novel global optimization method offers an efficient and effective solution for hyper-parameter tuning in deep learning.
• The enhanced auto-encoder model significantly improves automatic personality perception from speech.
• This work advances the application of deep learning in extracting nuanced human characteristics from vocal data.