EC3460 Introduction to Machine Learning for Signal Analytics

This course introduces basic concepts and tools needed to detect, analyze, model, and extract useful information from digital signals by finding patterns in data. It covers some of the fundamentals of machine learning as they apply in signal and information processing. The emphasis in the course is on practical engineering applications rather than theoretical derivations to give participants a broad understanding of the issues involved in the learning process. Supervised learning tools such as the Bayes estimator, neural networks and radial basis functions, support vector machines and kernel methods are presented. Unsupervised learning tools such as k-means and hierarchical clustering are discussed. Data transformation and dimensionality reduction are introduced. Performance measures designed to evaluate learning algorithms are introduced. Concepts presented are illustrated throughout the course via several application projects of specific interest to defense related communities. Application topics may include target/signal identification, channel equalization, speech/speaker recognition, image classification, blind source separation, power load forecasting, and others of current interest.

Prerequisite

Knowledge of probability and random variables (EC2010, or OS2080, or OA3101, or equivalent), linear systems (EC2410 or equivalent), linear algebra (MA2043 or equivalent), ability to program in MATLAB, or consent of instructor.

Lecture Hours

3

Lab Hours

2

Course Learning Outcomes

·       To learn what the learning problem is, and what its limitations are.

·       To be able to select a supervised or unsupervised approach to investigate the problem considered.

·       To be able to extract features characterizing class structures for classification applications.

·       To be able to design a Bayes classifier and apply it to a specific problem.

·       To be able to select a specific back-propagation neural network configuration to address a specific problem considered.

·       To be able to implement a back-propagation learning algorithm, train, and test the network.

·       To be able to apply principal component analysis to reduce the number of class features.

·       To be able to implement a radial basis function network to a classification or function approximation task.

·       To be able to apply the apply Kernel PCA to a classification problem.

·       To be able to select a given kernel type and apply it to a support vector machine algorithm.

·       To be able to apply an unsupervised learning method such as k-means or hierarchical clustering method to extract patterns present in signals.

·       To be able to select and apply a performance measure to evaluate and compare learning algorithms applied to a specific problem.

·       To be able to apply approaches introduced to specific applications of interest to defense related communities.