Multilayer perceptron architecture optimization using parallel computing techniques

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Science

Background:

• Multilayer-perceptron (MLP) neural networks are widely used for complex data analysis.
• Optimizing MLP architecture is crucial for improving model performance and accuracy.
• Neural network architecture parameters significantly influence prediction accuracy and error rates.

Purpose of the Study:

• To develop an optimization methodology for MLP neural networks.
• To evaluate the impact of hidden layers (HL), neurons per hidden layer (NHL), and activation function type (AF) on the sum of squares error (SSE).
• To analyze the efficiency of parallel computing in optimizing neural network architectures.

Main Methods:

• Utilized physicochemical and microbiological quality parameters of milk samples for training.
• Organized network architectures into three groups (G1, G2, G3) with one, two, and three hidden layers, respectively.
• Employed factorial experimental designs to generate 63, 2,187, and 50,049 combinations, varying NHL (3-27) and AF (tan-sig, log-sig, linear) using MATLAB 2015a and parallel computing.

Main Results:

• Number of hidden layers (HL) and neurons per hidden layer (NHL) significantly affect the sum of squares error (SSE).
• Activation function type (AF) demonstrates a significant effect on SSE starting from two hidden layers.
• An inverse relationship was observed between the number of processors and total optimization time across all groups.

Conclusions:

• HL, NHL, and AF are critical parameters for optimizing MLP neural networks.
• The developed methodology provides a systematic approach to identifying optimal neural network configurations.
• Parallel computing enhances the efficiency of neural network optimization, reducing overall computation time.