Programming instead of Neural Networks
Programming instead of Neural Networks
Overview
Exploring a pivot for the CS411 Digital Education project from AI-assisted learning of neural networks to programming concepts.
Upcoming Meeting Agenda
Discussion on the benefits, disadvantages, and trade-offs of switching the research domain to programming.
1. Potential Benefits
- Automated & Fine-Grained Evaluation: Code execution, unit test pass rates, syntax vs. logic error categorization provide objective benchmarks.
- Student Familiarity & Recruitment: Programming tasks generally align with a larger pool of EPFL/UNIL students compared to specialized neural network training intuition.
- Granular Telemetry: Easy to track keystroke cadence, compiler errors, AI prompt queries, and revision history.
2. Potential Disadvantages & Challenges
- Prior Experience Confound: Large variance in baseline coding skills among students may confound the measured impact of AI assistance.
- Scope & Open-Endedness: Debugging and code generation can become overly broad compared to constrained parameter tuning in a visual playground.
- Separating AI Help from Syntax Fixes: Risk of students relying on AI as a simple syntax autocomplete rather than engaging in conceptual learning.
Next Steps
- Align with team (Karo, Isna, Stergios) during the meeting on whether to pivot or maintain the NN playground setup.
- Define targeted programming concepts / exercises if pivot is approved.
- Review recruitment requirements (Michalis / Mauro student cohorts).
- Update experimental protocol (Group A vs Group B intervention).