这篇论文是2016 NIPS上的一篇论文。从标题就可以看出这篇论文的特点所在:Fast Localized Spectral Filtering。这篇论文的主要贡献在于:
- 改进了Spectral Networks,设计了 localized filter on graph,并且计算复杂度较低;
- 提出了一种较为有效的 pooling 策略。
Learning Fast Localized Spectral Filters
作者总结了将卷积推广到图上的三步骤:
- the design of localized convolutional filters on graphs;
- a graph coarsening (见《深入浅出图神经网络中》翻译成“图塌缩”)procedure that groups together similar vertices;
- a graph pooling operation that trades spatial resolution for higher filter resolution.
之前的 spectral networks,并不具备局部连接的特性,并且在图卷积层运算时,傅里叶变换的矩阵乘法的时间复杂度较高。
Polynomial parametrization for localized filters
针对这两个问题,作者借助泰勒展开去近似任意函数,从而将卷积核展开为:
g θ ( Λ ) = ∑ k = 0 K − 1 θ k Λ k g_\theta(\Lambda) = \sum_{k=0}^{K-1}{\theta_k \Lambda^k} gθ(Λ)=k=0∑K−1θkΛk
其中 θ ∈ R K \theta \in \mathbb{R}^K θ∈RK为可训练的多项式系数。如果将上式代入到 y = U g θ ( Λ ) U T x y=Ug_\theta(\Lambda)U^Tx y=Ugθ(Λ)UTx 可得到
y = ∑ k = 0 K θ k L k x y=\sum_{k=0}^{K}{\theta_kL^kx} y=k=0∑KθkLkx
设定 x ( k ) = L k x = L x ( k − 1 ) x^{(k)}=L^kx=Lx^{(k-1)} x(