Scaling Deep Learning Education in a College of Arts & Sciences

Artificial Intelligence systems powered by deep learning are changing how we work, communicate, and make decisions. If we want these technologies to serve society responsibly, tomorrow’s citizens need more than a superficial understanding of them, even if they are not computer science (CS) majors.

At the University of Michigan, the Department of Statistics took a bold step: we launched an accessible but rigorous deep learning course aimed at undergraduates in the College of Literature, Sciences, and the Arts (LSA). I was fortunate to be entrusted with leading this effort. Since its pilot offering in Winter 2022, the course has grown significantly and evolved in unexpected ways, from the topics we teach to the data students collect using their own sense of smell.

In this post, I share what we have learned in the past three years: the challenges we faced, the innovations we introduced, and what it takes to teach deep learning in a college of arts and sciences. I hope our experience will be useful to educators designing similar courses beyond traditional CS and engineering programs.

The Pilot

Designing a course that is both broadly accessible and intellectually rigorous was a challenge. Deep learning rests on a foundation of multivariable calculus, probability, linear algebra, and Python programming, a tall order for students outside engineering and CS.

To open the door to a broad audience, we pared down the enforced prerequisites to the essentials: single-variable calculus, an introductory programming course (in any language), and one course in statistics or probability. Because we were venturing into uncharted territory with such light prerequisites, we began with a small-scale pilot in Winter 2022, enrolling just 40 students. I focused my lectures on core concepts and mathematical intuitions, while the teaching assistants led lab sessions centered on programming and implementation.

We started with linear regression, a familiar concept for most students, but viewed through a deep learning lens. From there, we built up to the basic machine learning workflow: fitting models, adding regularization, evaluating performance, and then introducing fully connected neural networks.

But after a few weeks, I started to feel uneasy. Students seemed overwhelmed, not just by the math-heavy lectures, but also by the programming assignments, which took the form of mini-projects. It was not the students’ fault; my instruction had not fully adjusted to the realities of the reduced prerequisites.

The small class size made it easy to talk with students individually and understand where they were struggling. I paused the regular lectures and introduced a short bootcamp on linear algebra, paired with lab sessions on working with NumPy arrays and TensorFlow tensors in Python. The bootcamp helped fill key knowledge gaps and gave both the students and me a chance to regroup after the rocky start.

After the bootcamp, the course stabilized and students re-engaged. The small class size made it easy for them to give feedback and help steer the course in real time. I got to know nearly the entire class personally. Toward the end of the term, I introduced two more topics: convolutional neural networks (CNNs) for image data, and recurrent neural networks (RNNs) for time series. What had started as a near-disaster ended as a qualified success, made possible by the flexibility, feedback, and individual attention that a small pilot setting allowed.

Scaling Up to 100 Students

When I taught the course again in Winter 2023, enrollment easily surpassed 100 students, a clear sign of strong demand. This time, I introduced the linear algebra bootcamp at the beginning of the course rather than as a mid-semester adjustment. I also reworked how I taught the backpropagation algorithm, using only single-variable calculus and the chain rule. That approach eliminated the need for a separate multivariable calculus bootcamp.

By this point, the course structure had settled into four core modules: introductory material to ease students into deep learning, fully connected neural networks, convolutional networks for image data, and recurrent networks for time series.

After the release of ChatGPT in late 2022, students began asking about the transformer architecture, the “T” in GPT. In response, the Winter 2024 edition of the course replaced the RNN module with one on transformers. That same semester, I also received a teaching grant to support a mini-module on machine olfaction. The goal was twofold: to introduce students to an emerging research area and to give them a taste of the challenges involved in real-world data collection.

Introducing Machine Olfaction

Compared to vision, language, and audition, machine learning has had a relatively limited impact on olfaction, our sense of smell. But this is beginning to change. Google Brain researchers recently founded a company called Osmo, with the motto “Giving computers a sense of smell.” Deep learning pioneer and Turing Award recipient Geoffrey Hinton serves on its scientific advisory board. Its CEO, Alexander Wiltschko, a University of Michigan alumnus, graciously gave a guest lecture in my class. He introduced students to graph neural networks and their applications to molecular tasks, including predicting an odorant’s smell from its chemical structure.

Wiltschko’s lecture was paired with a hands-on session led by Michelle Krell Kydd, a sensory evaluation expert known locally as “the nose of Ann Arbor.” She guided students in attentively smelling various substances and recording their impressions. Later, a subset of students participated in a blind-smelling activity with 20 monomolecular odorants, labeling each one with descriptors such as “sweet” or “fruity.” The resulting data revealed both striking individual differences and shared perceptions. For instance, nearly every student labeled vanillin, the primary odorant in vanilla beans, as “vanilla.”

The introductory deep learning course, now listed as DATA SCI 315, has become a regular part of the statistics department’s undergraduate curriculum. I am excited to teach it again in Fall 2025 and plan to introduce students to generative models. I also hope to experiment with olfactory mixtures, which will challenge deep learning models to predict smell perception beyond monomolecular odorants.

Acknowledgments. I am grateful for the support of the College of Literature, Science, and the Arts (LSA) in the form of two NINI (New Initiatives/New Instruction) grants and an LSA Individual Award for Outstanding Contributions to Undergraduate Education. I also appreciate the support of my department leadership, especially Ed Ionides, Associate Chair for Undergraduate Studies, and Liza Levina, former Chair of the Department, for their bold vision for the course and their unwavering faith in my teaching abilities. I also thank my faculty colleagues Yixin Wang and Jeff Regier for sharing their valuable experience in refining the course format and delivery. I am also deeply grateful to all my students and teaching assistants who patiently suffered through the early iterations and gave feedback that helped improve the later versions of the course.

Ambuj Tewari is a professor of statistics and, by courtesy, computer science at the University of Michigan. His research and teaching focus on machine learning and artificial intelligence, with applications in psychiatry and chemistry.