Laurae (version 0.0.0.9001)

loss_LKL_grad: Laurae's Kullback-Leibler Error (gradient function)

Description

This function computes the Laurae's Kullback-Leibler loss gradient per value provided preds and labels values.

Usage

loss_LKL_grad(y_pred, y_true)

Arguments

y_pred

The predictions.

y_true

The labels.

Value

The gradient of the Laurae's Kullback-Leibler Error per value.

Details

This loss function is strictly positive, therefore defined in \]0, +Inf\[. It penalizes lower values more heavily, and as such is a good fit for typical problems requiring fine tuning when undercommitting on the predictions. Compared to Laurae's Poisson loss function, Laurae's Kullback-Leibler loss has much higher loss. This loss function is experimental. Loss Formula : \((y_true - y_pred) * log(y_true / y_pred)\) Gradient Formula : \(-((y_true - y_pred)/y_pred + log(y_true) - log(y_pred))\) Hessian Formula : \(((y_true - y_pred)/y_pred + 2)/y_pred\)