Sample complexity | EPFL Graph Search

The sample complexity of a machine learning algorithm represents the number of training-samples that it needs in order to successfully learn a target function. More precisely, the sample complexity is the number of training-samples that we need to supply to the algorithm, so that the function returned by the algorithm is within an arbitrarily small error of the best possible function, with probability arbitrarily close to 1. There are two variants of sample complexity:

The weak variant fixes a particular input-output distribution;
The strong variant takes the worst-case sample complexity over all input-output distributions.

The No free lunch theorem, discussed below, proves that, in general, the strong sample complexity is infinite, i.e. that there is no algorithm that can learn the globally-optimal target function using a finite number of training samples. However, if we are only interested in a particular class of target functions (e.g, only linear functions) then the sample

A Universal Sampling Method for Reconstructing Signals with Simple Fourier Transforms

Mikhail Kapralov, Amir Zandieh

Reconstructing continuous signals based on a small number of discrete samples is a fundamental problem across science and engineering. We are often interested in signals with "simple" Fourier structure - e.g., those involving frequencies within a bounded range, a small number of frequencies, or a few blocks of frequencies i.e., bandlimited, sparse, and multiband signals, respectively. More broadly, any prior knowledge on a signal's Fourier power spectrum can constrain its complexity. Intuitively, signals with more highly constrained Fourier structure require fewer samples to reconstruct. We formalize this intuition by showing that, roughly, a continuous signal from a given class can be approximately reconstructed using a number of samples proportional to the statistical dimension of the allowed power spectrum of that class. We prove that, in nearly all settings, this natural measure tightly characterizes the sample complexity of signal reconstruction. Surprisingly, we also show that, up to log factors, a universal non-uniform sampling strategy can achieve this optimal complexity for any class of signals. We present an efficient and general algorithm for recovering a signal from the samples taken. For bandlimited and sparse signals, our method matches the state-of-the-art, while providing the the first computationally and sample efficient solution to a broader range of problems, including multiband signal reconstruction and Gaussian process regression tasks in one dimension. Our work is based on a novel connection between randomized linear algebra and the problem of reconstructing signals with constrained Fourier structure. We extend tools based on statistical leverage score sampling and column-based matrix reconstruction to the approximation of continuous linear operators that arise in the signal reconstruction problem. We believe these extensions are of independent interest and serve as a foundation for tackling a broad range of continuous time problems using randomized methods.

ASSOC COMPUTING MACHINERY2019

Fourier Sampling in Signal Processing and Numerical Linear Algebra

Amir Zandieh

This thesis focuses on developing efficient algorithmic tools for processing large datasets. In many modern data analysis tasks, the sheer volume of available datasets far outstrips our abilities to process them. This scenario commonly arises in tasks including parameter tuning of machine learning models (e.g., Google Vizier) and training neural networks. These tasks often require solving numerical linear algebraic problems on large matrices, making the classical primitives prohibitively expensive. Hence, there is a crucial need to efficiently compress the available datasets. In other settings, even collecting the input dataset is extremely expensive, making it vital to design optimal data sampling strategies. This is common in applications such as MRI acquisition and spectrum sensing. The fundamental questions above are often dual to each other, and hence can be addressed using the same set of core techniques. Indeed, exploiting structured Fourier sparsity is a recurring source of efficiency in this thesis, leading to both fast numerical linear algebra methods and sample efficient data acquisition schemes. One of the main results that we present in this thesis is the first Sublinear-time Model-based Sparse FFT algorithm that achieves a nearly optimal sample complexity for recovery of every signal whose Fourier transform is well approximated by a small number of blocks (e.g., such structure is common in spectrum sensing). Our method matches in sublinear time the result of Baraniuk et. al. (2010), which started the field of model-based compressed sensing. Another highlight of this thesis includes the first Dimension-independent Sparse FFT algorithm that, computes the Fourier transform of a sparse signal in sublinear runtime in any dimension. This is the first algorithm that just like the FFT of Cooley and Tukey is dimension independent and avoids the curse of dimensionality inherent to all previously known techniques. Finally, we give a Universal Sampling Scheme for the reconstruction of structured Fourier signals from continuous measurements. Our approach matches the classical results of Slepian, Pollak, and Landau (1960s) on the reconstruction of bandlimited signals via Prolate Spheroidal Wave Functions and extends these results to a wide class of Fourier structure types. Besides having classical applications in signal processing and data analysis, Fourier techniques have been at the core of many machine learning tasks such as Kernel Matrix Approximation. The second half of this thesis is dedicated to finding compressed and low-rank representations of kernel matrices, which are the primary means of computation with large kernel matrices in machine learning. We build on Fourier techniques and achieve spectral approximation guarantees to the Gaussian kernel using an optimal number of samples, significantly improving upon the classical Random Fourier Features of Rahimi and Recht (2008). Finally, we present a nearly-optimal Oblivious Subspace Embedding for high-degree Polynomial kernels which leads to nearly-optimal embeddings of the high-dimensional Gaussian kernel. This is the first result that does not suffer from an exponential loss in the degree of the polynomial kernel or the dimension of the input point set, providing exponential improvements over the prior works, including the TensorSketch of Pagh (2013) and application of the celebrated Fast Multipole Method of Greengard and Rokhlin (1986) to the kernel approximation problem.

EPFL2020

Efficiently Learning Fourier Sparse Set Functions

Mikhail Kapralov, Andreas Krause, Amir Zandieh

Learning set functions is a key challenge arising in many domains, ranging from sketching graphs to black-box optimization with discrete parameters. In this paper we consider the problem of efficiently learning set functions that are defined over a ground set of size n and that are sparse (say k-sparse) in the Fourier domain. This is a wide class, that includes graph and hypergraph cut functions, decision trees and more. Our central contribution is the first algorithm that allows learning functions whose Fourier support only contains low degree (say degree d = o(n)) polynomials using O(kd log n) sample complexity and runtime O(kn log(2) k log n log d). This implies that sparse graphs with k edges can, for the first time, be learned from O(k log n) observations of cut values and in linear time in the number of vertices. Our algorithm can also efficiently learn (sums of) decision trees of small depth. The algorithm exploits techniques from the sparse Fourier transform literature and is easily implementable. Lastly, we also develop an efficient robust version of our algorithm and prove l(2)/l(2) approximation guarantees without any statistical assumptions on the noise.

NEURAL INFORMATION PROCESSING SYSTEMS (NIPS)2019