FidelityLoss — Gnostic Fidelity Loss¶
The FidelityLoss rewards predictions that align with the target-centered concept manifold by maximizing mean(sech(2θ)).
Overview¶
For each residual-centered coordinate θ = (y_pred - y_true) / S, the loss evaluates the fidelity response fi = sech(2θ) and maximizes its mean. The objective lives in [0, 1), reaches optimum 1 when predictions match targets exactly, and approaches 0 as fidelity collapses.
Key Features¶
- Custom autograd kernel (
FidelityLossFunction) for efficient gradients - Scale parameter
Smay be numeric or"auto"(batch-derived) y_truesupplies fixed concept center and never receives gradients- Gradient clipping for numerical stability
Parameters¶
| Parameter | Type | Default | Description |
|---|---|---|---|
S |
float | str |
1 |
Scale parameter; use "auto" for local batch scale |
name |
str | None |
None |
Optional display name |
verbose |
bool |
False |
Verbose logging |
Example Usage¶
import numpy as np
from machinegnostics.magnet import FidelityLoss, Tensor
y_pred = Tensor(np.array([[0.0], [0.0]]), requires_grad=True)
y_true = Tensor(np.array([[0.0], [0.0]]))
print(float(FidelityLoss()(y_pred, y_true))) # 1.0
Author: Nirmal Parmar