Use of Parallel Explanatory Models to Enhance Transparency of Neural Network Configurations for Cell Degradation...

Area of Science:

• Machine Learning
• Signal Processing
• Telecommunications

Background:

• Recurrent Neural Networks (RNNs) accurately detect cellular network radio signal degradations.
• Adding layers to RNNs unexpectedly diminished accuracy gains, prompting further investigation.

Purpose of the Study:

• To build a parallel model for understanding the internal operations of neural networks (NNs) that process sequential inputs.
• To investigate the diminishing accuracy gains observed with increased RNN layers.

Main Methods:

• Developed a parallel model applicable to input domains representable by Gaussian mixtures.
• Analyzed RNN processing using a probability density function (pdf) perspective.
• Validated the model against RNN processing stages and output predictions.

Main Results:

• Demonstrated how each RNN layer transforms input distributions to enhance detection accuracy.
• Identified a side effect limiting accuracy improvements in RNNs.
• Explained the reasons behind RNN performance limitations.

Conclusions:

• The parallel model provides insights into RNN internal operations and accuracy limitations.
• Findings offer valuable guidance for designing future RNNs and similar NNs.
• Understanding the pdf transformation is key to optimizing sequential data processing in NNs.