Skip to content

CFGD (Conformable)

Implements CFGD (Conformable Fractional Gradient Descent), a local fractional optimizer that rescales the gradient by a conformable factor.

Classical fractional gradient methods built on the Caputo or Riemann–Liouville operators are nonlocal: they accumulate a weighted history of past gradients, which is costly to maintain during training. CFGD instead uses the conformable fractional derivative of Khalil et al., \(T_\alpha f(\theta) = \theta^{1-\alpha}\,f'(\theta)\), which is local and reduces to a simple power-law rescaling of the ordinary derivative. The update therefore costs essentially the same as plain gradient descent while the fractional order \(\alpha \in (0,1]\) modulates the effective step per coordinate.

Applied coordinate-wise to the loss, the conformable gradient rescales each component \(g_t\) by \(|\theta_t|^{1-\alpha}\), recovering standard gradient descent at \(\alpha = 1\). A variable-order variant (vCFGD) lets the order grow toward \(1\) during training, \(\alpha_t \uparrow 1\), giving stronger fractional modulation early and gradient-descent–like behavior asymptotically.

\[ \begin{aligned} \theta_{t+1} &= \theta_t - \eta\,|\theta_t|^{\,1-\alpha_t}\, g_t, \\ g_t &= \nabla_\theta \mathcal{L}(\theta_t), \\ \alpha_t &\uparrow 1 \quad (\text{vCFGD}). \end{aligned} \]

where \(\theta\) are the parameters, \(\eta\) the learning rate, \(g_t\) the gradient, \(\alpha_t \in (0,1]\) the (possibly time-varying) conformable order, and \(|\theta_t|^{1-\alpha_t}\) the conformable rescaling factor applied per coordinate; setting \(\alpha_t = 1\) recovers ordinary gradient descent.

Reference: Hayman Thabet, "Conformable fractional gradient descent: A local optimizer for neural network training", Journal of Computational and Applied Mathematics 488 (2026). https://doi.org/10.1016/j.cam.2026.117842


Back to the Canon