Graduation Semester and Year

Fall 2024

Language

English

Document Type

Thesis

Degree Name

Master of Science in Electrical Engineering

Department

Electrical Engineering

First Advisor

Michael T Manry

Abstract

Interpreting multi-layer perceptron (MLP) classifier outputs as posterior probabilities is a well-established practice in machine learning and is supported in the literature. However, several authors point out that MLP outputs are very poor estimates of the posterior probabilities. This is demonstrated for classifiers with and without nonlinear output activation. Achieving this reliability depends on key factors such as model complexity, sufficient training data availability, and optimization techniques' effectiveness. In practice, these requirements are not met, resulting in suboptimal probability estimates. Our approach introduces an innovative method based on the softmax output. The method aim to refine MLP discriminants into more reliable posterior probability estimates. We validate our framework through extensive experiments on diverse real-world datasets, comparing its performance against Naive Bayesian classifiers. Our results show substantial improvements in the accuracy of probability estimates, marking a significant advance in probabilistic inference for neural network outputs.

Keywords

Neural Network, feedforward, Multiple Optimal Learning factors, Scalable, Machine Learning, Posterior Probability, Probability Estimation, Activation Functions

Disciplines

Controls and Control Theory | Signal Processing

License

Creative Commons Attribution 4.0 International License
This work is licensed under a Creative Commons Attribution 4.0 International License.

Available for download on Monday, June 16, 2025

Share

COinS