The AI Skills Employers Want: Gonzaga’s Data 575 Class

How do tech professionals stay ahead of the curve in the rapidly evolving world of Artificial Intelligence (AI) and Large Language Models (LLMs)? Gonzaga University’s M.S. in Data Science and AI program addresses that challenge with the class “Data 575: From Natural Language Processing to Large Language Models”. Designed by Joseph Dumoulin, the course bridges data science and machine learning with human language processing, equipping students with practical skills to stand out in interviews and excel in their careers.

The course dives into the mechanics of how machines interpret human speech to text.

“The idea of this class is really to talk about how language in general, formal or informal natural languages are able to be automated in this way. It’s become a pretty important aspect of what we talk about when we talk about AI these days. The key is the natural language processing part. How does that work? What is the process that it goes through to take in language and then spit out language and to give the data scientist tools for understanding deeply what’s going on when they encounter an LLM,” said Dumoulin.

Dumoulin brings decades of industry experience to the classroom. He began his career in the mid-1980’s as a computer programmer developing factory maintenance software, followed by five years running his own company focused on software solutions for small engineering companies. By the early 2000s, he began working with chatbots, eventually transitioning to Aon insurance in 2018 to apply Natural Language Processing (NLP) to legal documents and patents.

His background directly shapes the course curriculum, which opens with the 50-year history of natural language to help students build a strong foundation.

I’ve been doing this since 2002, so that’s been a huge part of my professional career. I am very familiar with all these methods that nobody uses anymore,” said Dumoulin.

Understanding methods, such as Recurrent Neural Networks that predate modern LLMs, is essential. It helps students grasp how the simpler architecture of LLMs aids in its ability to consume vast amounts of data quickly. Building on this knowledge, students engage in hands-on practice using real-world scenarios to practice fine-tuning LLMs.

“I want to try and introduce the use of agent coding methods early on so that students can get an idea of how to use multiple agents and how to tell the agent what to do so it doesn’t go off and do something nutty, which they do. If you learn how to build the constraints, then you can get much better performance.”

Dumoulin continued, “We will learn how to fine tune open source LLMs to solve specific NLP tasks, and we will use coding agents to build our environment for coding with a large language model. It will tell us what code to do, and we’ll tell it what we want to do with fine tuning and where our data lives and it will write the code and then run the fine tuning.”

Fine tuning is an in-demand skill that organizations use to tune models to their own domain.

“Knowing how to do fine tuning is a big plus when you work in industry. At the same time, teaching how to use coding tools that are available in LLMs is really helpful when you are being interviewed for a new data science or programming position. It really requires a ton of sophistication with the tools.”

Dumoulin continued, “Fine tuning is a comprehensive skill for Machine Learning applications. You have to do this all the time. Knowing how to find models that have the pre-trained aspects you need and then learning how to build your “fine tuning” framework and then introduce training, debug the training, make sure it works and then get a final result that you can test. That is the stuff I think has a lot of durability.”

To help students keep pace with these rapidly evolving skills, Dumoulin incorporates current research alongside foundational texts. Students analyze recently published papers covering emerging techniques that have not yet reached standard literature, such as Flash Attention, which is a method for optimizing transformers to work more efficiently in large networks.

At the same time, the curriculum pulls from core standards using the seminal textbook Speech and Language Processing by Daniel Jurafsky and James H. Martin. Aligning with Gonzaga’s broader mission, ethics are also deeply woven into the coursework, with weekly discussions focusing on current news and emerging ethical considerations in AI.

When summarizing the ultimate impact of the course, Dumoulin noted its broad value for modern software developers and data scientists alike.

“It’s valid for anybody who is a programmer. Mostly students get to learn how natural language processing neural networks have evolved to get to this point. Learn how the fine tuning of natural language models works and learn how to use large language model agents to do coding.”

Data 575 was built by an industry professional with decades of knowledge to help ensure Gonzaga graduates are uniquely equipped to navigate and lead the future of artificial intelligence.