InfidelityLoss — Gnostic Infidelity Loss¶
Penalize inverse fidelity so off-concept predictions become expensive.
Overview¶
InfidelityLoss minimizes the batch mean of the infidelity characteristic fj = cosh(2θ) with θ = (y_pred - y_true) / S. Because fj is bounded in [1, ∞), the loss emphasizes poor concept alignment more aggressively than fidelity-only objectives.
Value range: [1, ∞) (minimum 1 at perfect alignment).
Key Features¶
- Custom autograd kernel (
InfidelityLossFunction) - Scale parameter
S(numeric or"auto") - Sharp penalty for low-fidelity predictions
Parameters¶
| Parameter | Type | Default | Description |
|---|---|---|---|
S |
float | str |
1 |
Scale parameter; "auto" for local batch scale |
name |
str | None |
None |
Optional display name |
verbose |
bool |
False |
Verbose logging |
Attributes¶
S: Scale parameterS_local: Resolved local scale
Methods¶
forward(y_pred, y_true)¶
Return mean inverse-fidelity objective (mean(cosh(2θ))).
Example Usage¶
import numpy as np
from machinegnostics.magnet import InfidelityLoss, Tensor
y_pred = Tensor(np.zeros((2, 1)), requires_grad=True)
y_true = Tensor(np.zeros((2, 1)))
print(float(InfidelityLoss()(y_pred, y_true))) # 1.0
Author: Nirmal Parmar