ml
Convergence Checker
Absolute, relative, and plateau convergence detection — EMA smoothing, exponential decay fit (L∞ + L₀e^{-λt}), extrapolated convergence point.
Loss Sequence
Convergence Criterion
Loss — raw / EMA / exponential fit
—— raw—— EMA--- exp fit
Convergence
Steps20
Final loss0.445
|Δloss| last0.002
Rel Δloss last0.00447
Converged at stepNot yet
Exponential Decay Fit
L∞ (floor)0.445
L₀ (amplitude)9.323
λ (decay rate)0.5367
Half-life (steps)1.29
Extrapolated conv (step)12.7
Not yet converged under absolute criterion (ε = 0.01)
EMA: s_t = α·loss_t + (1-α)·s_{t-1}
Fit: loss(t) = L∞ + L₀·e^(-λt) via log-linearization
Extrapolated convergence: t* = ln(L₀/ε) / λ
Fit: loss(t) = L∞ + L₀·e^(-λt) via log-linearization
Extrapolated convergence: t* = ln(L₀/ε) / λ