arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 1702.08489cs.LGcs.CCstat.ML
收藏

神经网络的深度分离

Depth Separation for Neural Networks

Google Brain(谷歌大脑)

Amit Daniely

摘要:

令$f:\mathbb{S}^{d-1}\times \mathbb{S}^{d-1}\to\mathbb{S}$为$g:[-1,1]\to \mathbb{R}$的形式为$f(\mathbf{x},\mathbf{x}') = g(\langle\mathbf{x},\mathbf{x}'\rangle)$的函数。我们给出了一个简单的证明,表明多尺寸深度2神经网络(指数)有界的权重不能近似$f$时,$g$不能近似的低次多项式。而且,对于许多$g$,如$g(x)=\sin(πd^3x)$,神经元的数量必须是$2^{Ω\left(d\log(d)\right)}$。此外,该结果也适用于w.r.t.\ $\mathbb{S}^{d-1}\times \mathbb{S}^{d-1}$上的均匀分布。由于上述形式的许多函数可以很好地近似于具有多边形有界权重的多边形深度三网络,因此这在深度二和深度三网络之间建立了分离。$\mathbb{S}^{d-1}\times \mathbb{S}^{d-1}$上的均匀分布。

英文摘要:

Let $f:\mathbb{S}^{d-1}\times \mathbb{S}^{d-1}\to\mathbb{S}$ be a function of the form $f(\mathbf{x},\mathbf{x}') = g(\langle\mathbf{x},\mathbf{x}'\rangle)$ for $g:[-1,1]\to \mathbb{R}$. We give a simple proof that shows that poly-size depth two neural networks with (exponentially) bounded weights cannot approximate $f$ whenever $g$ cannot be approximated by a low degree polynomial. Moreover, for many $g$'s, such as $g(x)=\sin(πd^3x)$, the number of neurons must be $2^{Ω\left(d\log(d)\right)}$. Furthermore, the result holds w.r.t.\ the uniform distribution on $\mathbb{S}^{d-1}\times \mathbb{S}^{d-1}$. As many functions of the above form can be well approximated by poly-size depth three networks with poly-bounded weights, this establishes a separation between depth two and depth three networks w.r.t.\ the uniform distribution on $\mathbb{S}^{d-1}\times \mathbb{S}^{d-1}$.

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。

↑