Skip to content

Caputo-Type FOGD (Deep BP)

Implements Caputo-Type FOGD (Deep BP), fractional-order gradient descent for deep backpropagation networks using the Caputo derivative.

The method replaces the integer-order gradient with a Caputo fractional derivative of order \(v \in (0,1)\) of the loss with respect to each weight. Applying the Caputo operator to the power-law form \(\theta^{1-v}\) that arises from the backpropagated error yields a closed-form fractional gradient scaled by the Gamma function, which is then used in the ordinary descent step. The fractional order interpolates between memory-laden updates and the classical first-order rule recovered as \(v \to 1\).

\[ \begin{aligned} D^{v}_{\theta} E &= \frac{g_t \, \theta^{1-v}}{\Gamma(2 - v)}, \\ \theta_{t+1} &= \theta_t - \eta \, D^{v}_{\theta} E. \end{aligned} \]

where \(\theta\) is a weight, \(g_t = \delta^{l+1} a^{l}\) is the standard integer-order error gradient backpropagated to that weight, \(v \in (0,1)\) is the fractional order, \(\Gamma(\cdot)\) is the Gamma function, and \(\eta > 0\) is the learning rate. With \(L_2\) regularization of strength \(\lambda\), the fractional gradient gains the term \(\lambda \theta \, 2^{1-v} / \Gamma(3 - v)\).

Reference: Y. Chen, G. Zhao, "A Caputo-type fractional-order gradient descent learning of deep BP neural networks", 2019 IEEE 3rd Advanced Information Management, Communicates, Electronic and Automation Control Conference (IMCEC), pp. 546-550, 2019. https://doi.org/10.1109/IMCEC46724.2019.8984089


Back to the Canon