Skip to content

Q-Newton

Implements Q-Newton, Newton's gradient descent with hybrid quantum-classical scheduling of the Hessian inversion.

Q-Newton trains neural networks with the full Newton step, preconditioning the gradient by the inverse Hessian. Its contribution is a scheduler that, at each step, estimates the cost of inverting the Hessian on a classical LU solver versus a quantum linear-systems solver and dispatches to whichever is cheaper; the parameter-update rule itself is the standard regularized Newton iteration. The Hessian is optionally regularized as \(H + \epsilon I\) to improve conditioning before inversion.

\[ \begin{aligned} \Delta\theta_t &= -\,(H_t + \epsilon I)^{-1}\, g_t, \\ \theta_{t+1} &= \theta_t + \beta\, \Delta\theta_t. \end{aligned} \]

where \(g_t = \nabla_\theta \mathcal{J}(\theta_t)\) is the gradient of the loss, \(H_t = \nabla_\theta^2 \mathcal{J}(\theta_t)\) is its Hessian, \(\beta\) is the (fixed) learning rate, \(\epsilon\) is the regularization term, and \(I\) is the identity matrix.

Reference: Pingzhi Li, Junyu Liu, Hanrui Wang, Tianlong Chen, "Hybrid Quantum-Classical Scheduling for Accelerating Neural Network Training with Newton's Gradient Descent", arXiv 2024. https://arxiv.org/abs/2405.00252


Back to the Canon