In December, I had an eye-opening conversation with Prof. Rahul Mangharam at the University of Pennsylvania about the challenges of teaching embedded AI to undergraduate students. That discussion prompted me to reflect on how we, as instructors, design curriculum that balances accessibility with depth. What follows is a set of thoughts on how embedded AI can be introduced in ways that invite students in (without overwhelming them or lowering expectations).
Overview
Embedded AI (a subset of “Edge AI” and often called “TinyML”) has become increasingly difficult to ignore in engineering education. It promises industry relevance, cross-disciplinary learning, and compelling projects. It also presents a real pedagogical challenge. Embedded AI sits at the intersection of hardware, firmware, signal processing, and machine learning. Each of those domains is demanding on its own. When we introduce them all at once, we risk overwhelming students.
When educators ask how embedded AI should be taught, my answer is usually: it depends on where students are in the curriculum and what learning outcomes we’re targeting. In practice, I think the most effective programs deliberately combine two complementary approaches: a bottom-up approach that emphasizes depth and synthesis, and a top-down approach that emphasizes scaffolding, motivation, and early engagement.
Bottom-Up: Synthesis
The bottom-up approach treats embedded AI as a capstone topic rather than an entry point. This aligns well with traditional engineering curricula and with pedagogical models that emphasize mastery learning and conceptual dependency.
In this model, students encounter embedded AI only after they’ve built substantial foundations, including the following:
- Electronics and sensing fundamentals – Courses covering sensors, analog signal conditioning, sampling, quantization, and noise provide the grounding needed to understand where embedded ML data actually comes from and why data quality matters.
- Microcontroller architecture and embedded systems – Students develop a mental model of memory hierarchies, peripherals, interrupts, timing, and resource constraints, allowing them to reason about what is reasonable to run on a microcontroller and what problems they can solve.
- Embedded software development in C/C++ – Prior experience writing, debugging, and maintaining embedded firmware helps students appreciate determinism, memory management, and the practical realities of deploying code to constrained devices.
- Real-time systems concepts – Exposure to scheduling, latency, and timing guarantees prepares students to understand why inference time, responsiveness, and predictability matter in embedded AI applications.
- Signal processing and DSP – Often serving as the conceptual bridge, topics such as sampling theory, filtering, FFTs, and feature extraction help students work with time series data as inputs to ML models.
- Machine learning fundamentals – With the above context in place, machine learning concepts like classification versus regression, training versus inference, and model evaluation can be taught in a way that is grounded in real constraints such as limited memory, latency budgets, and power consumption. This often includes working with Python and one (or more) of the popular frameworks, like PyTorch.
From a pedagogical perspective, this approach treats embedded AI as a synthesis course. Here, students are integrating knowledge from multiple prior courses. When successful, this produces graduates who can reason deeply about trade-offs, debug performance issues, and recognize when machine learning might not be the best solution for a problem.
The limitation is not rigor but accessibility. Bottom-up embedded AI courses work best as upper-level electives, capstones, or graduate courses. For earlier students, the delayed payoff can be discouraging. If their first meaningful encounter with embedded AI comes late in the program, many will opt out before they ever reach it.
This is where a complementary approach becomes valuable.
Top-Down: Scaffolding and Orientation
The top-down approach reframes embedded AI as an orientation experience rather than a mastery experience. Rather than waiting for students to accumulate all prerequisites, it introduces embedded AI early using deliberate scaffolding.
Here, the instructional goal is not theoretical completeness but high-level understanding with early wins. Students interact with real sensors and microcontrollers early. They collect data, train simple models, and deploy inference quickly using tools like Edge Impulse. The emphasis is on answering the question “what does embedded AI do, and why does it matter?”
This approach aligns naturally with project-based learning. Short, tightly scoped projects give students concrete success criteria and visible outcomes. Higher-level tooling and abstractions reduce cognitive load, allowing students to focus on core ideas such as data quality, system constraints, and end-to-end behavior rather than low-level implementation details.
For sophomores or motivated freshmen, this kind of exposure is often transformative. Students leave with intuition rather than mastery. They understand that running inference on a microcontroller is fundamentally different from cloud ML. They see why latency, power, and memory shape design decisions. Just as importantly, they begin to see why future courses in signal processing, probability, and embedded systems will matter.
From a pedagogical standpoint, this is classic instructional scaffolding. While complexity and theory are deferred to later in the student’s learning journey, such an approach can hook a student’s interest early to keep them wanting to learn more.
Avoiding the False Dichotomy
Problems arise when we treat these approaches as mutually exclusive or expect one to achieve the learning outcomes of the other.
A top-down course should not be expected to produce embedded AI experts. Its role is to build intuition, interest, and context. Likewise, a bottom-up course is not an effective recruitment mechanism for students who have not yet decided whether this field is for them.
Strong programs design a learning progression rather than a single course. Early exposure courses use projects and scaffolding to spark curiosity. Core courses build the mathematical and systems foundations. Advanced electives then revisit embedded AI in a bottom-up, integrative way (i.e. with significantly deeper technical expectations).
This mirrors a spiral curriculum, where students revisit the same domain multiple times with increasing sophistication. What begins as intuition eventually becomes analysis. What begins as tooling eventually becomes theory.
Final Thoughts
Embedded AI is quickly becoming a core competency rather than a niche specialization. If we want students to engage with it meaningfully, we need to be intentional about sequencing. Early, scaffolded exposure builds confidence and curiosity. Later, rigorous synthesis builds capability. When aligned with modern pedagogical practices like project-based learning and spiral curricula, these approaches reinforce rather than compete with each other.
Teaching embedded AI well isn’t about lowering standards. It’s about meeting students where they are and designing pathways that allow them to grow into the full complexity of the field.