Goals

Primary Goal

Practice inventing energy-efficient learning algorithms for simple non-trivial learning tasks.

Sub-goals

1. Solve sparse parity with a neural network (>90% accuracy on 3-bit parity with 17 noise bits)
2. Estimate energy using Average Reuse Distance as proxy metric
3. Improve energy usage by modifying the learning algorithm
4. Share tips and AI prompting strategies with the group

Research Questions

1. Can modern AI make a learning algorithm to solve simple learning tasks?
2. Can it improve (memory) energy usage of those algorithms?
3. What prompting strategies and approaches are useful for this?

Success Criteria

• Training + evaluation runs in <1 second
• Measurable ARD improvement over standard backprop baseline
• Document the process and prompting strategies used