93
44
v1v2v3 (latest)

Efficient active learning of sparse halfspaces with arbitrary bounded noise

Abstract

We study active learning of homogeneous ss-sparse halfspaces in Rd\mathbb{R}^d under the setting where the unlabeled data distribution is isotropic log-concave and each label is flipped with probability at most η\eta for a parameter η[0,12)\eta \in \big[0, \frac12\big), known as the bounded noise. Even in the presence of mild label noise, i.e. η\eta is a small constant, this is a challenging problem and only recently have label complexity bounds of the form O~(spolylog(d,1ϵ))\tilde{O}\big(s \cdot \mathrm{polylog}(d, \frac{1}{\epsilon})\big) been established in [Zhang, 2018] for computationally efficient algorithms. In contrast, under high levels of label noise, the label complexity bounds achieved by computationally efficient algorithms are much worse: the best known result of [Awasthi et al., 2016] provides a computationally efficient algorithm with label complexity O~((slndϵ)2poly(1/(12η)))\tilde{O}\big((\frac{s \ln d}{\epsilon})^{2^{\mathrm{poly}(1/(1-2\eta))}} \big), which is label-efficient only when the noise rate η\eta is a fixed constant. In this work, we substantially improve on it by designing a polynomial time algorithm for active learning of ss-sparse halfspaces, with a label complexity of O~(s(12η)4polylog(d,1ϵ))\tilde{O}\big(\frac{s}{(1-2\eta)^4} \mathrm{polylog} (d, \frac 1 \epsilon) \big). This is the first efficient algorithm with label complexity polynomial in 112η\frac{1}{1-2\eta} in this setting, which is label-efficient even for η\eta arbitrarily close to 12\frac12. Our active learning algorithm and its theoretical guarantees also immediately translate to new state-of-the-art label and sample complexity results for full-dimensional active and passive halfspace learning under arbitrary bounded noise. The key insight of our algorithm and analysis is a new interpretation of online learning regret inequalities, which may be of independent interest.

View on arXiv
Comments on this paper