arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2308.02001cs.LGmath.OC
收藏

具有平滑激活函数的两层神经网络的记忆容量

Memory capacity of two layer neural networks with smooth activations

University of British Columbia(不列颠哥伦比亚大学)

Liam Madden, Christos Thrampoulidis

AI总结:

本文研究两层神经网络的记忆容量,针对实解析且满足次数条件的平滑激活函数,建立了 $\lfloor md/2\rfloor$ 的下界并证明其近似最优,通过推导雅可比矩阵的通用秩推广了已有结论。

摘要:

确定具有 $m$ 个隐藏神经元和输入维度 $d$(即 $md+2m$ 个总可训练参数)的两层神经网络的记忆容量是一个基本的机器学习问题,该容量指的是网络能够记忆的通用数据的最大规模。对于在某点处实解析且若限制为该点处的多项式时具有足够高次数的激活函数,我们建立了 $\lfloor md/2\rfloor$ 的下界,并证明了其最优性误差在约 $2$ 倍以内。所有实际的激活函数,如 sigmoids、Heaviside 和 rectified linear unit (ReLU),在某点处都是实解析的。此外,次数条件很宽松,例如,如果激活函数为 $x^k$,则要求 $\binom{k+d-1}{d-1}\ge n$。类似的先前结果仅限于 Heaviside 和 ReLU 激活函数——我们的结果几乎涵盖了所有其他情况。为了分析一般激活函数,我们推导了网络雅可比矩阵的精确通用秩,该秩可以用 Hadamard 幂和 Khatri-Rao 乘积表示。我们的分析扩展了关于 Hadamard 幂秩的经典线性代数事实。总体而言,我们的方法不同于先前关于记忆容量的工作,并有望扩展到更深层的模型和其他架构。

英文摘要:

Determining the memory capacity of two layer neural networks with $m$ hidden neurons and input dimension $d$ (i.e., $md+2m$ total trainable parameters), which refers to the largest size of general data the network can memorize, is a fundamental machine learning question. For activations that are real analytic at a point and, if restricting to a polynomial there, have sufficiently high degree, we establish a lower bound of $\lfloor md/2\rfloor$ and optimality up to a factor of approximately $2$. All practical activations, such as sigmoids, Heaviside, and the rectified linear unit (ReLU), are real analytic at a point. Furthermore, the degree condition is mild, requiring, for example, that $\binom{k+d-1}{d-1}\ge n$ if the activation is $x^k$. Analogous prior results were limited to Heaviside and ReLU activations -- our result covers almost everything else. In order to analyze general activations, we derive the precise generic rank of the network's Jacobian, which can be written in terms of Hadamard powers and the Khatri-Rao product. Our analysis extends classical linear algebraic facts about the rank of Hadamard powers. Overall, our approach differs from prior works on memory capacity and holds promise for extending to deeper models and other architectures.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑