CSC4651/5651 Deep Learning In Signal Processing: Project Milestone 1: Topic Selection and References

This page describes the first of six project milestones.

For your term project, teams of 2 are recommended. Teams of 3 are allowed if the project has sufficient scope. You may work individually if you prefer. Make only 1 submission per team in Canvas (from any student on the team) and clearly identify the team members on the first page.

All projects must implement and evaluate a DNN (deep neural network). MATLAB’s Deep Learning Toolbox is recommended, but you may choose another appropriate tool/API if you prefer; TensorFlow (including Keras) and PyTorch are acceptable—please check with the professor before choosing other APIs. audioFlux may be useful if you choose a Python API.

You might expand the scope, especially if you are building on an existing solution; the scope may be adjusted as you research the topic. For example, you may choose to focus on transfer learning, model pruning, quantization, and/or deployment.

For this milestone, you must choose a general topic (e.g., product classification from images using transfer learning, estimating user age from voice recording) and identify a minimum of 3 technical resources; having 5-10 references may be reasonable depending on the quality and quantity of what you find. A minimum of 1 (and ideally at least 2 or 3) of your resources should be conference or journal papers published by IEEE, ACM or similar, high-quality sources. Textbooks, blogs, etc., are acceptable as background and supplemental references.

You should look at abstracts and outlines of your sources with the goal of identifying promising materials (for the next assignment you’ll take a deeper dive into your sources). You want your sources to enable you to answer the following questions.

• How has the same or a similar problem been solved with DNNs and perhaps without DNNs?
• What are the challenges in solving this problem?
• Is an appropriate dataset freely available? (Some are installed in Rosie in common locations, which makes it easy for any user to load them; we can add more upon request.)
• Is it appropriate to evaluate your solution using a basic error function like MSE or binary cross-entropy, or should additional measures of performance (e.g., quality metrics accepted in the industry) be considered?

For each reference, include a short paragraph explaining why it should be useful.

Some places to get started:

• IEEEXplore—the MSOE Library has access to most IEEE publications
• Google Scholar
• MSOE Library Guides
  • Computer Science Resources
  • Datasets and API Resources