CSC 5651 - Deep Learning in Signal Processing

4 lecture hours 0 lab hours 4 credits

Course Description
This elective course provides an overview of deep learning methods and models as used in digital signal processing (DSP), including key DSP concepts that appear in and adjacent to such models in both real-time and off-line applications. The course begins with basics of creating and evaluating deep learning models and then proceeds to alternate between DSP and deep learning topics. Deep learning structures such as convolutional layers, recurrent networks, dropout, and autoencoders are covered. DSP topics including frequency response, the role of convolution, and spectrograms are covered with an emphasis on how they support deep learning models. Examples of audio and image applications (time and space as independent variables of a potentially multidimensional sampled signal) are included throughout, supporting further work in video processing, medical image processing, etc. A variety of current models are studied throughout the term. Topics of student interest are addressed by special lecture topics and course projects. Laboratory exercises include several weeks of guided exercises and culminate with a term project.

Graduate students enrolled in the course will prepare critical summaries of published research papers in addition to meeting the requirements of the undergraduate course. Milestones of these summaries will include regular reviews and feedback from the faculty member and presentations to the class. (prereq: (MTH 5810 or MTH 2340 or MTH 2130), (CSC 5610 or ELE 3320 or CSC 3310 or CSC 3310 or CSC 2621)) (quarter system prereq: (MA 383 or BE 2200), (CS 2040 or CS 3210 or EE 3221))

Course Learning Outcomes
Upon successful completion of this course, the student will be able to:

Describe state of the art deep learning structures for addressing signal processing problems
Articulate the calculations done during backpropagation and explain various challenges that may occur (exploding gradients, uncontrollable parameters, etc.)
Evaluate proposed solutions and interpret published results using standard metrics such as accuracy, precision, recall, top-N accuracy, etc.
Break down modern deep learning architectures into their components and characterize the function of those components
Modify existing network structures and evaluate the impact on model performance for new applications
Appraise a data set in terms of quantity, quality, and balance, and suggest appropriate mitigations of data set issues
Apply the spectrogram as a foundation for signal detection, classification, and enhancement
Differentiate between commonly used objective and subjective metrics and critique their use in backpropagation and model evaluation
Relate a signal processing view of convolution to a deep learning view through the concepts of time-invariance, linearity, and feature extraction
Evaluate recent research results in the field
Distill, summarize in writing, and present recent research results to peers with basic ML and DSP knowledge

Prerequisites by Topic

Linear algebra and/or multivariable calculus
Basic software design, ideally in the context of data science, numeric applications, or AI

Course Topics

Training pipelines
Loss functions including perceptual losses
Confusion matrices and performance metrics
Convolutional layers of various dimensions used on both time series and time-frequency representations of data
Mitigation of overfitting including dropout and batchnorm
Principal components analysis and autoencoders
Various types of RNNs
Common network architectures
Frequency response
Discrete Fourier transforms
Spectrograms, windowing, and perfect reconstruction

Laboratory Topics

Basic DNN training
Signal representation / transfer learning
Model pruning
Hyperparameter optimization
Project
Presentations

Coordinator
Dr. Eric Durant

Add to Portfolio.

Close Window