A postdoctoral fellow of our group Dr. Aku Jaakko Kammonen participated in the recently concluded Conference on Computational Science and Engineering (CSE23) and presented a talk on Adaptive random Fourier features based on Metropolis sampling. The Conference held at Amsterdam, Netherlands, between February 26 - March 3, 2023.
Abstract:
The supervised learning problem to approximate a function $f:\mathbb R^d\to\mathbb R$ by a neural network approximation $\mathbb{R}^d\ni x\mapsto\sum_{k=1}^K\hat\beta_k e^{{\mathrm{i}}\omega_k\cdot x}$ with one hidden layer is studied as a random Fourier features algorithm.
Here the mean square loss problem can be solved easily, since it is convex in the amplitude parameters $\hat\beta_k$ given a density $p:\mathbb R^d\to [0,\infty)$ for independent frequencies $\omega_k$. It is also well known that the corresponding generalization error is bounded by $K^{-1}\Vert \vert \hat f\vert ^2/((2\pi)^dp)\Vert _{L^1(\mathbb R^d)} $, where $\hat f$ is the Fourier transform of $f$. In my talk I will first show how the constant $\Vert \vert \hat f\vert ^2/((2\pi)^dp)\Vert _{L^1(\mathbb R^d)}$ can be minimized by optimally choosing the density $p$ and then how to approximately sample from this density, only using the data and certain adaptive Metropolis steps. I will also show results with other activation functions.
[1] Kammonen, Aku and Kiessling, Jonas and Plechac, Petr and Sandberg, Mattias and Szepessy, Anders. Adaptive random Fourier features with Metropolis sampling. Foundations of Data Science, 2020.
[2] Kammonen, Aku and Kiessling, Jonas and Plechac, Petr and Sandberg, Mattias and Szepessy, Anders and Tempone, Raul. Smaller generalization error derived for a deep residual neural network compared with shallow networks. IMA Journal of Numerical Analysis, 2022.