Mono-Forward algorithm offers local learning alternative to backpropagation

Researchers have introduced Mono-Forward (MF), a new algorithm designed to improve upon the Forward-Forward (FF) method for training deep neural networks. MF maintains the local learning and reduced memory footprint of FF while replacing its contrastive goodness objective with a standard cross-entropy loss. This modification allows MF to achieve competitive or superior performance compared to FF and even backpropagation on certain tasks, such as MLP-Mixers on PathMNIST, while using significantly less memory. AI

影响 Introduces a more memory-efficient and competitive alternative to backpropagation for deep learning training.

排序理由 This is a research paper introducing a new algorithm for training neural networks. [lever_c_demoted from research: ic=1 ai=1.0]

在 arXiv cs.LG 阅读 →

• backpropagation
• MLP-Mixers
• Mono-Forward
• PathMNIST

• 论文
• 其他

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →