---
id: 20260422-T0-12
title: "BASIS：新型神经网络反向传播优化方案"
title_en: "BASIS Optimizes Neural Network Backpropagation"
url: https://ai.daily.yangsir.net/daily/20260422-T0-12
issue_date: 2026-04-22
publish_date: 2026-04-21T04:00:00.000Z
category: research
source_name: "arXiv cs.LG (ML)"
source_url: https://arxiv.org/abs/2604.16324
---

# BASIS：新型神经网络反向传播优化方案

arXiv论文提出BASIS方法，通过平衡激活剪影和不标量不变量解决'鬼反向传播'问题。传统反向传播的内存需求与网络深度呈线性关系，形成计算瓶颈。该方案显著降低内存使用，特别适用于处理长序列和深层网络的训练场景，可提升大模型训练效率。

## English Version

**BASIS Optimizes Neural Network Backpropagation**

arXiv paper introduces BASIS, a method that solves 'ghost backpropagation' through balanced activation sketching and invariant scalars. Traditional backpropagation memory scales linearly with network depth, creating bottlenecks. This approach significantly reduces memory usage, particularly beneficial for training large models on long sequences.

---

**来源**：[arXiv cs.LG (ML)](https://arxiv.org/abs/2604.16324)

**详情页**：https://ai.daily.yangsir.net/daily/20260422-T0-12

---

*智语观潮 · Daily — https://ai.daily.yangsir.net/llms.txt*