Can your own update schedule reach a low loss?
Specify a learning-rate schedule and compare repeated updates from two starting points.
For each rate, update w ← w − rate × 2(w−target). Retain every step and loss.
Cut these cards apart. Use the original example first, then let a partner supply a changed case. Keep sealed/final cards with the teacher until the learner commits a rule.
Start: -2
Learning rate schedule: 0.1,0.1,0.1,0.1
My prediction and reason:
Start: 0
Learning rate schedule: 1.1,1.1,1.1,1.1
My prediction and reason:
Start: 8
Learning rate schedule: 0.5,0.2,0.1
My prediction and reason:
This convex quadratic is a controlled teaching case, not a neural training benchmark.