Skip to main content

Programming Neural Networks for Novice Learners

Project Overview

Project concept refined during team discussion (Group: Kyoto baby). Rather than abandoning neural networks, the project will investigate programming of neural networks tailored for novice/beginner programmers (using neural networks as the concrete applied use case).


Core Concept & Pedagogical Approach

  • The Synthesis (Programming + NNs):
    • Neural networks are the use case, while code-based implementation is the learning medium.
    • Targets students who recently completed an introductory coding course (e.g., Éléments de programmation—familiar with basic syntax, functions, and basic OOP/classes, but still novices).
    • Dual Objective: Measure how learners improve both their algorithmic problem-solving skills and their conceptual understanding of neural network mechanics.
  • Interface & Scaffolding Strategy:
    • Interactive web-based IDE / notebook style (single cell execution / stepwise progression).
    • Move beyond simple click-based GUI playgrounds by giving learners structured code snippets to modify.
    • Scaffolded complexity: Start with sensible default parameters and gradually introduce optional arguments, logic branches, and hyperparameters.
  • Evaluation Metrics:
    • Focus on conceptual grasp, application ability, and knowledge transfer to new problem contexts.
    • Avoid superficial metrics (e.g., raw lines of code or CPU execution time); focus on logic comprehension and error resolution.

Project Timeline & Milestone Roadmap

Target / Date Objective Details
This Week 1-Page Proposal Draft Outline hypotheses, independent/dependent variables, and study protocol for Mauro & Dominique (TA).
Lunch Today Recruitment Alignment Stergios asking Yash to contact 2nd-year Bachelor students. Master's students excluded (avoiding high baseline expertise).
October 8 (Milestone 1) Official Idea Validation Finalize recruitment forms, validate participant pipeline (~20 novices), and secure TA approval.
By Week 8 (Milestone 2) Data Collection & Initial Findings Complete ~20 participant sessions (~30 total testing hours, ~2 full days per researcher). Have preliminary results synthesized.
Milestone 3 Final Analysis & Reporting ~1.5 months dedicated to deep analytics, drawing conclusions, and drafting the project paper/Kyoto submission.

Action Items Checklist

  • Stergios to speak with Yash over lunch regarding UNIL/EPFL 2nd-year Bachelor recruitment.
  • Draft 1-page proposal summary (independent/dependent variables, web IDE task flow).
  • Send 1-pager to Dominique (TA) for early pre-validation ahead of Milestone 1.
  • Finalize pre-screening test instrument (objective test to filter genuine novices).
  • Milestone 1 deadline: October 8th.

Meeting Notes: Experimental Design & UI Specifications

  • Experimental Design & Cohort Balancing:
    • Participants: ~20 Bachelor students (strictly non-Master to preserve novice baseline).
    • Pre-Test / Pre-Quiz: Administered prior to grouping to assess baseline programming skill and balance proficiency evenly across test groups.
    • Conditions: 10 with AI assistant vs. 10 control (or testing specific feedback modalities, e.g., immediate vs. Socratic/delayed guidance, avoiding multi-variable confounding).
  • Interface & Interaction Architecture:
    • Split-Screen Layout: Code editor on one side, dynamic real-time visualization of the neural network on the other.
    • Visual Feedback: Visual architecture updates live or step-by-step as students write/modify code blocks, reinforcing mental models.
    • AI Assistant Boundaries (Strict): No code auto-completion (no Copilot-style generation) for any group. The LLM only provides guided hints, Socratic questions, and high-level suggestions so students must write all implementation code themselves.
  • Protocol & Evaluation:
    • Playground Phase: ~30 minutes of structured exploratory practice.
    • Post-Test Transfer Task: Implement an NN for a new, slightly more advanced use case.
    • Metrics: Algorithmic correctness/logic evaluation (assessed manually/systematically), completion time, and conceptual transfer.

Implementation Scope & Team Workstream Division

  • Curriculum & Scope Boundaries:
    • Aligned strictly with TensorFlow Playground introductory concepts (layers, iterations/epochs, basic loss functions). Avoid unnecessary complexity beyond this baseline.
    • Stergios to share reference materials / brief summaries explaining loss functions for team onboarding.
  • Pre-Test Metrics & Assessment Instrument:
    • Lead: Isna (coordinating directly with Mauro) to define the pre-test skill measurement metrics.
  • Platform Implementation & Prototyping:
    • Lead: Stergios.
    • Timeline: Working prototype targeted by end of this week.
    • Staging Host: neuralnet.stergios.ch (dedicated domain to be deployed for the live study).
  • 1-Page Proposal Drafting:
    • Section responsibilities assigned; phrasing and narrative being finalized.

Target Publication: IJCAI 2027 (Special Track on Human-Centred AI)

  • Venue: 36th International Joint Conference on Artificial Intelligence (IJCAI 2027)
  • Location: Kyoto, Japan (August 7–13, 2027) — matches team name Kyoto baby
  • Submission Deadline: ~Mid-January 2027 (Abstract ~Jan 12, Full paper ~Jan 19)
  • Format: IJCAI LaTeX template, 7 pages content + 2 pages references (max 9 pages total), double-blind.
  • Scope & Requirements for Human-Centred AI (HAI) Track:
    • Must emphasize the human-centric AI system design, interaction paradigm, and usability/cognitive scaffolding, rather than pure pedagogy alone.
    • Rigorous empirical evaluation / user study (the ~20 EPFL Bachelor novice cohort comparing AI scaffolding vs. control).
    • Clear operationalization of human-AI collaboration, error correction, and cognitive load/knowledge transfer.

Study Design Evolution & Proposal (Discussion Update - Isna)

  • Motivation / Participant Feasibility:

    • Concern raised that a full neural network programming case study is likely too advanced for target novice participants.
    • Suggestion to pivot back toward supporting specific core conceptual understanding (e.g., gradient descent, bias, or hidden layers) via interactive visualization tools rather than end-to-end network implementation.
  • Pedagogical Framing (ICAP Framework):

    • Visualization: Targets Passive or Active engagement depending on how learners interact with visual components.
    • Coding (Optional / Trade-off): Could potentially target Constructive learning, but risks adding excessive cognitive load. Note: coding is not mandatory if the primary objective is testing AI role dynamics.
    • AI Support (Interactive Engagement): Exploring the effect of AI agent persona/role:
      • AI as Tutor (expert guidance, hints/scaffolding) vs.
      • AI as Peer (co-learner with similar knowledge baseline).
  • Immediate Decisions for Next Group Meeting (Mon/Tue):

    1. Define and agree upon the exact learning goal / target concept to teach (e.g., gradient descent mechanics vs. another novel concept).
    2. Decide whether coding is retained as a constructive condition or replaced entirely by interactive visualization + AI roles.
    3. Finalize study conditions ahead of Milestone 1 validation.

Team Discussion: Topic Selection & Experimental Feasibility (Meeting Notes)

    Experimental Constraints & Group Sizing:

      Session Length: Target learning task must be strictly doable within ~1 hour (learning + evaluation). Statistical Power & Groups: With ~20 total participants, keep to 2 groups (10 with AI assistant vs. 10 without). Avoid 2x2 factorials (e.g., Code vs. Visual × AI vs. No-AI = 4 groups of 5), which dilutes statistical validity.

      Evaluated Concept Options:

        Option A: Gradient Descent (Focused NN Subtopic)
          Pros: Fits the neural network umbrella; complementary to standard course material. Logistics: Stergios is lecturing and running the lab next Monday (2:00–4:00 PM lecture, 4:00–6:00 PM lab) while Michalis is on vacation. Could pilot with the 50 students in class, but because Milestone 1 feedback won't be returned yet, this may serve only as a pilot pre-experiment requiring 20 separate participants later. Considerations: Must ensure the digital tool provides distinct value beyond basic high-school calculus/derivatives. Option B: Text Analytics & Dimensionality Reduction
            Context: Covered later in Michalis' syllabus (Oct 19 / Oct 26) with TF-IDF, text embeddings, and dimensionality reduction. Option C: Recursion ("Zero-to-Hero" Algorithmic Learning)
              Pros: Quintessential programming hurdle for novices; self-contained, conceptual, and highly evaluable in a 1-hour intervention. Target Audience: 1st/2nd-year Bachelor students with basic programming knowledge (recursion covered in Week 2). Eliminates complex mathematical baggage while testing genuine algorithmic transfer.

              Immediate Next Step:

                Final decision needed today on the core topic (Gradient Descent vs. Recursion) to finalize the 1-page proposal for Milestone 1 (Oct 8).