BriefGPT.xyz
Sep, 2019
轻度过参数化神经网络能够高效地记忆训练数据
Mildly Overparametrized Neural Nets can Memorize Training Data Efficiently
HTML
PDF
Rong Ge, Runzhe Wang, Haoyu Zhao
TL;DR
本文研究发现，为了在过量参数化情况下实现 100％ 的准确率，神经网络可以训练出能够完美记忆训练数据的能力，即使在神经元数量要远小于训练样本数量的情况下依旧适用。
Abstract
It has been observed \citep{zhang2016understanding} that
deep neural networks
can
memorize
: they achieve 100\% accuracy on
training data
.
→