Revisiting Agnostic PAC Learning
- URL: http://arxiv.org/abs/2407.19777v1
- Date: Mon, 29 Jul 2024 08:20:49 GMT
- Title: Revisiting Agnostic PAC Learning
- Authors: Steve Hanneke, Kasper Green Larsen, Nikita Zhivotovskiy,
- Abstract summary: PAC learning, dating back to Valiant'84 and Vapnik and Chervonenkis'64,'74, is a classic model for studying supervised learning.
Empirical Risk Minimization (ERM) is a natural learning algorithm, where one simply outputs the hypothesis from $mathcalH$ making the fewest mistakes on the training data.
We revisit PAC learning and first show that ERM is in fact sub-optimal if we treat the performance of the best hypothesis, denoted $tau:=Pr_mathcalD[hstar_math
- Score: 30.67561230812141
- License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
- Abstract: PAC learning, dating back to Valiant'84 and Vapnik and Chervonenkis'64,'74, is a classic model for studying supervised learning. In the agnostic setting, we have access to a hypothesis set $\mathcal{H}$ and a training set of labeled samples $(x_1,y_1),\dots,(x_n,y_n) \in \mathcal{X} \times \{-1,1\}$ drawn i.i.d. from an unknown distribution $\mathcal{D}$. The goal is to produce a classifier $h : \mathcal{X} \to \{-1,1\}$ that is competitive with the hypothesis $h^\star_{\mathcal{D}} \in \mathcal{H}$ having the least probability of mispredicting the label $y$ of a new sample $(x,y)\sim \mathcal{D}$. Empirical Risk Minimization (ERM) is a natural learning algorithm, where one simply outputs the hypothesis from $\mathcal{H}$ making the fewest mistakes on the training data. This simple algorithm is known to have an optimal error in terms of the VC-dimension of $\mathcal{H}$ and the number of samples $n$. In this work, we revisit agnostic PAC learning and first show that ERM is in fact sub-optimal if we treat the performance of the best hypothesis, denoted $\tau:=\Pr_{\mathcal{D}}[h^\star_{\mathcal{D}}(x) \neq y]$, as a parameter. Concretely we show that ERM, and any other proper learning algorithm, is sub-optimal by a $\sqrt{\ln(1/\tau)}$ factor. We then complement this lower bound with the first learning algorithm achieving an optimal error for nearly the full range of $\tau$. Our algorithm introduces several new ideas that we hope may find further applications in learning theory.
Related papers
- Learning a Single Neuron Robustly to Distributional Shifts and Adversarial Label Noise [38.551072383777594]
We study the problem of learning a single neuron with respect to the $L2$ loss in the presence of adversarial distribution shifts.
A new algorithm is developed to approximate the vector vector squared loss with respect to the worst distribution that is in the $chi2$divergence to the $mathcalp_0$.
arXiv Detail & Related papers (2024-11-11T03:43:52Z) - Sample and Computationally Efficient Robust Learning of Gaussian Single-Index Models [37.42736399673992]
A single-index model (SIM) is a function of the form $sigma(mathbfwast cdot mathbfx)$, where $sigma: mathbbR to mathbbR$ is a known link function and $mathbfwast$ is a hidden unit vector.
We show that a proper learner attains $L2$-error of $O(mathrmOPT)+epsilon$, where $
arXiv Detail & Related papers (2024-11-08T17:10:38Z) - Deterministic Apple Tasting [2.4554686192257424]
We provide the first widely-applicable deterministic apple tasting learner.
We prove a trichotomy stating that every class $mathcalH$ must be either easy, hard, or unlearnable.
Our upper bound is based on a deterministic algorithm for learning from expert advice with apple tasting feedback.
arXiv Detail & Related papers (2024-10-14T11:54:46Z) - On Optimal Learning Under Targeted Data Poisoning [48.907813854832206]
In this work we aim to characterize the smallest achievable error $epsilon=epsilon(eta)$ by the learner in the presence of such an adversary.
Remarkably, we show that the upper bound can be attained by a deterministic learner.
arXiv Detail & Related papers (2022-10-06T06:49:48Z) - Cryptographic Hardness of Learning Halfspaces with Massart Noise [59.8587499110224]
We study the complexity of PAC learning halfspaces in the presence of Massart noise.
We show that no-time Massart halfspace learners can achieve error better than $Omega(eta)$, even if the optimal 0-1 error is small.
arXiv Detail & Related papers (2022-07-28T17:50:53Z) - Threshold Phenomena in Learning Halfspaces with Massart Noise [56.01192577666607]
We study the problem of PAC learning halfspaces on $mathbbRd$ with Massart noise under Gaussian marginals.
Our results qualitatively characterize the complexity of learning halfspaces in the Massart model.
arXiv Detail & Related papers (2021-08-19T16:16:48Z) - Hardness of Learning Halfspaces with Massart Noise [56.98280399449707]
We study the complexity of PAC learning halfspaces in the presence of Massart (bounded) noise.
We show that there an exponential gap between the information-theoretically optimal error and the best error that can be achieved by a SQ algorithm.
arXiv Detail & Related papers (2020-12-17T16:43:11Z) - Model-Free Reinforcement Learning: from Clipped Pseudo-Regret to Sample
Complexity [59.34067736545355]
Given an MDP with $S$ states, $A$ actions, the discount factor $gamma in (0,1)$, and an approximation threshold $epsilon > 0$, we provide a model-free algorithm to learn an $epsilon$-optimal policy.
For small enough $epsilon$, we show an improved algorithm with sample complexity.
arXiv Detail & Related papers (2020-06-06T13:34:41Z) - Agnostic Learning of a Single Neuron with Gradient Descent [92.7662890047311]
We consider the problem of learning the best-fitting single neuron as measured by the expected square loss.
For the ReLU activation, our population risk guarantee is $O(mathsfOPT1/2)+epsilon$.
For the ReLU activation, our population risk guarantee is $O(mathsfOPT1/2)+epsilon$.
arXiv Detail & Related papers (2020-05-29T07:20:35Z)
This list is automatically generated from the titles and abstracts of the papers in this site.
This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.