AI

Hybrid Quantum-Classical Scheduling for Accelerating Neural Network Training with Newton’s Gradient Descent

View a PDF file from the paper entitled Q -NewUTON: Classic hybrid scheduling to accelerate the nerve network training with the descent of Newton, by Pingzhi Li and 3 other authors

PDF HTML (experimental) view

a summary:deep learning improvement techniques often lead first -class gradient methodologies, such as SGD. However, nervous network training can greatly benefit from the characteristics of rapid rapprochement to improve the second ranking. Newton’s GD stands out in this category, by reassessing the gradient using the reverse hessian. However, one of its main bottlenecks is the reflection of the matrix, which takes a long time at $ o (n^3) with weak expansion.

The reflection of the matrix can be translated into solving a series of linear equations. Given that the quantum linear solutions (QLSAS), can work on the basis of the principles of quantum overlap and tangle, can work within a $ \ {polylog} (N) (N) (N) $, it provides a promising approach with a huge acceleration. Specifically, one of the latest QLSAs shows a complex limitation of $ O (D \ cdot \ Kappa \ Log (N \ CDOT \ Kappa/\ Epsilon) $, depending on: {Size ~ $ N $, $ ~ $ \ $ $, Reeror \ Epsilon $, Qual. However, this also means that its potential assembly can be hindered by some real estate (i.e. $ \ Kappa $ and $ D $ D).

We suggest Q -NewUTON, which is a classic hybrid quantity to accelerate the nervous network training with GD Newton. Q -NEWUTON uses a simplified scheduling unit coordinating between quantum and classic linear solvents, by estimating and reducing $ \ Kappa $ and $ D $ D $ D for quantitative solution.

Our evaluation shows the possibility to significantly reduce Q -NewUTON to the total training time compared to commonly used altoings such as SGD. We assume a future scenario where the time of quantum machines is reduced, and may be achieved by AttoseConds. Our evaluation determines a promising ambitious goal for the development of quantum computing.

The application date

From: Bingzi Lee [view email]
[v1]

Tuesday, 30 April 2024 23:55:03 UTC (957 KB)
[v2]

Friday, 11 April 2025 19:21:02 UTC (417 KB)
[v3]

Tuesday, April 29, 2025 06:14:40 UTC (418 KB)

Don’t miss more hot News like this! AI/" target="_blank" rel="noopener">Click here to discover the latest in AI news!

2025-04-30 04:00:00

Related Articles

Back to top button