Multi-Epoch Separation Under Composition

Investigating whether function composition creates optimization barriers that separate learning epochs for component vs. composed functions.

~1.0
Separation Ratio (all classes)
99.3%
MSE Increase (relu_2layer)
5
Function Families Tested
4
Curriculum Strategies
6
Composition Depths

Component vs. Composed MSE by Function Family

Composition Depth Scaling (Final MSE)

Curriculum Strategy Comparison (poly_deg2)

Curriculum Strategy Comparison (relu_2layer)

Curvature Sensitivity: Separation Ratio vs Scale

Training Curves: Component f vs Composed (poly_deg2)

Function Family Summary (Exp 1)

Function FamilyMean f MSEMean g MSEMean Comp MSEMSE GapSep. Ratio
Polynomial deg20.07840.08710.095021.3%1.000
Polynomial deg30.15760.17230.182515.9%1.000
ReLU 2-layer0.09400.09660.187499.3%1.000
ReLU 3-layer0.14610.13980.242065.6%1.000
Piecewise linear0.67390.62580.869429.0%1.000

Curriculum Strategy Results

Strategypoly_deg2 MSErelu_2layer MSEpoly_deg2 Epochsrelu_2layer Epochs
Direct0.09880.169160.060.0
Sequential0.14650.259060.060.0
Warmstart0.07070.134157.660.0
Progressive0.11880.188960.060.0