Explainable AI w/Python
Today, I would like to share an AI SoTL paper entitled, “ Explainable Artificial Intelligence with MicroPython: Lightweight Neural Networks for Students’ Deeper Learning” by Klinkhammer (2025) (https://doi.org/10.3991/ijet.v20i04.55567 ).
The author investigates how embedding neural networks into resource-constrained microcontrollers and coding them directly in MicroPython can deepen students’ conceptual grasp of AI fundamentals. Utilizing the open-source framework AI-ANNE which facilitates transferring models from high-level libraries like TensorFlow and Keras to lightweight MicroPython implementations the research compares three instructional conditions:
traditional high-level libraries,
MicroPython in a standard environment, and
MicroPython on microcontrollers.
Statistical analysis using Tukey’s HSD test revealed that students who implemented neural networks from scratch in MicroPython demonstrated significantly higher learning gains than peers using TensorFlow and Keras, regardless of whether microcontrollers were involved. While the use of microcontrollers did not further elevate learning outcomes beyond the MicroPython environment, the hands-on, low-abstraction coding approach clearly supports greater transparency and explainability in AI education by demystifying core operations such as matrix multiplication, weight adjustment, and activation functions.
These findings underscore the pedagogical value of reducing abstraction in AI instruction to strengthen understanding of model mechanics and reinforce trust, transparency, and usability in explainable AI contexts.
From a learning-theory standpoint, this study aligns closely with constructionist and experiential learning perspectives, which emphasize learning through making, debugging, and iterating on tangible artifacts. Papert’s (1980) constructionism argues that learners develop deeper understanding when they actively build computational objects that externalize their thinking precisely what occurs when students implement neural networks line-by-line in MicroPython rather than relying on opaque, high-level libraries. Similarly, Dewey’s (1938) theory of experiential learning underscores the role of direct interaction with materials and processes in fostering meaning; coding lightweight neural networks on constrained devices situates learning in concrete problem-solving contexts that demand reflection and inquiry. Kolb’s (1984) experiential learning cycle is also evident, as students move iteratively between concrete coding experiences, reflective observation (examining model behavior), abstract conceptualization (understanding neural network mechanics), and active experimentation (modifying parameters and logic). By intentionally lowering abstraction, the approach supports epistemic agency and sense-making, reinforcing the idea that explainable AI learning is most powerful when students engage with the process of computation rather than merely its outputs.
Reference
Klinkhammer, D. (2025). Explainable AI with MicroPython: Lightweight Neural Networks for students’ deeper learning. International Journal of Emerging Technologies in Learning (iJET), 20(4), 72–80.