Testably Learning Polynomial Threshold Functions

Rubinfeld & Vasilyan recently introduced the framework of testable learning as an extension of the classical agnostic model. It relaxes distributional assumptions which are difficult to verify by conditions that can be checked efficiently by a tester. The tester has to accept whenever the data truly satisfies the original assumptions, and the learner has to succeed whenever the tester accepts. We focus on the setting where the tester has to accept standard Gaussian data. There, it is known that basic concept classes such as halfspaces can be learned testably with the same time complexity as in the (distribution-specific) agnostic model. In this work, we ask whether there is a price to pay for testably learning more complex concept classes. In particular, we consider polynomial threshold functions (PTFs), which naturally generalize halfspaces. We show that PTFs of arbitrary constant degree can be testably learned up to excess error $\varepsilon > 0$ in time $n^{\mathrm{poly}(1/\varepsilon)}$. This qualitatively matches the best known guarantees in the agnostic model. Our results build on a connection between testable learning and fooling. In particular, we show that distributions that approximately match at least $\mathrm{poly}(1/\varepsilon)$ moments of the standard Gaussian fool constant-degree PTFs (up to error $\varepsilon$). As a secondary result, we prove that a direct approach to show testable learning (without fooling), which was successfully used for halfspaces, cannot work for PTFs.

Paper

Similar papers

Peer review

Reviewer 9MWZ7/10 · confidence 2/52024-06-29

Summary

This paper studies testably learning an n-dimensional Polynomial Threshold function using a reduction proved in a previous work called fooling. The authors give an analysis of a construction of fooling multilinear PTF and then further for fooling arbitrary PTF. The paper completes itself with proof that push-forward cannot learn PTF and that fooling is the best they can have.

Strengths

The topic is very interesting, as there is a line of work on general testable learning and testable learning in specific classes like halfspace. Testable learning PTF is the natural next step. Though the authors use the technique of reducing testable learning to fooling from previous work, it seems the analysis of the construction of a fooling for multilinear PTF is also an important technique. As the authors mentioned, the previous fooling construction [GKK23] only works for degree 2 PTF, and the construction from [Kane19] needs some careful analysis to work for PTF with more than constant degree. The paper is complete as they also show fooling is necessary as another more direct approach cannot testably learn PTF.

Weaknesses

The last part of the paper shows that push-forward cannot learn PTF seems very interesting, but it seems hard for me to determine how significant the contribution is from the analysis of fooling multilinear PTF.

Questions

I am wondering what are the original techniques the authors want to emphasize. Like using Taylor expansion to bound additional error terms or reducing PTF to multilinear PTF or anything else? I am wondering if the authors can emphasize a bit more.

Rating

7

Confidence

2

Soundness

3

Presentation

3

Contribution

2

Limitations

N/A

Reviewer zio47/10 · confidence 4/52024-07-12

Summary

Background: Agnostic learning is a well-studied framework that models learning when no function is some hypothesis class F describes the data perfectly. Specifically, the agnostic learning framework requires the learning algorithm to give a hypothesis whose classification error is at most opt+$\epsilon$, where opt is the best prediction error among all hypotheses in the class F. Almost all existing agnostic learning algorithms are distribution-specific, i.e. they assume the examples are drawn from some distribution, for example a Gaussian distribution. A distribution-specific agnostic learning algorithm lacks in reliability, because it is allowed to output an extremely poor classifier if the examples do not come from e.g. Gaussian (or some other assumed distribution). Yet, fully eliminating such assumptions is shown to be impossible for many basic function classes (based on well-established cryptographic assumptions). Testable learning is a framework that aims to mitigate the above mentioned limitation, by allowing the algorithm to abstain on a specific dataset if the examples do not come from the assumed distribution. Overall, this allows a user to be confident that the classifier indeed has error of at most opt+$\epsilon$, as required by the agnostic learning framework. Testable learning has been a focus of many works in recent years (see the paper for references). The paper studies testable learning of polynomial threshold functions (PTFs) under the Gaussian distribution. I.e. the function class F considered in this work consists of functions of the form sign$(p(x))$, where p is a degree-$d$ polynomial. The work in n dimensions, with accuracy parameter $\epsilon$, the paper gives an algorithm for testable learning of constant-degree PTFs with a run-time of $n^{poly(1/\epsilon)}$}. The paper is based on the moment-matching framework of [Gollakota, Klivans, Kothari ‘23], and shows that the direct approach used in [Vasilyan, Rubinfeld ‘23] to handle linear threshold functions cannot be extended to PTFs. In order to apply the moment-matching framework of [Gollakota, Klivans, Kothari ‘23], the paper shows that PTFs are “fooled” by distributions whose low-degree moments are close to Gaussian moments. To do this, the paper expands on the approach of [Kane ‘11] that proves a less general statement that PTFs are “fooled” by distributions for which the marginal of every k coordinates equals to the k-dimensional Gaussian. The proof first first considers multilinear PTFs, and then reduces the case of general PTFs to that of multilinear PTFs.

Strengths

- Polynomial threshold functions are an extremely well-studied class of hypotheses that has been the focus of many works in learning theory (including in works that appeared in NeurIPS). - Previously no testable learning algorithms were known even for degree-2 polynomial threshold functions. - The run-time of $n^{poly(1/\epsilon)}$ qualitatively matches the best run-time for the agnostic learning of polynomial threshold functions. For example, existing hardness results preclude run-times such as $poly(n/\epsilon)$ or $n^{polylog(1/\epsilon)}$. - Studying polynomial threshold functions naturally extends previous works that study linear threshold functions.

Weaknesses

-The run-time dependence of the algorithm on the degree d of the PTF can conceivably be sub-optimal. As explained on page 12, the run-time is $(n \epsilon)^{O_d(\epsilon^{-4d7^d})}$, whereas it is conceivable that this run-time could potentially be improved in the future to $(n \epsilon)^{poly(d/\epsilon)}$

Questions

The paper mentions that the analysis does not require to extend the Carbery-Wright inequality to distributions whose low-degree moments match those of the Gaussian. Could you give some more high-level intuitive explanation for why this is the case?

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

I think that the limitations are discussed adequately

Reviewer 2r1G6/10 · confidence 2/52024-07-13

Summary

This paper studied the problem of testably learning polynomial threshold functions (PTFs). The authors aimed to answer the question of whether PTFs are qualitatively harder to learn in the testably learning model, compared to the agnostic learning model. The authors answered the question in the negative, showing that the degree-d PTFs can be testably learned up to $\epsilon$ with respect to the standard gaussian in time and sample complexity $n^{poly(1/\epsilon)}$, which qualitatively matches the $n^{O(d^2/\epsilon^4)}$ sample complexity of agnostic learning degree-d PTFs. To prove the above result, the authors linked testable learning with distribution fooling, building upon previous results on polynomial approximations. The authors also showed that it is impossible to testably learn PTFs with the techniques from [RV23].

Strengths

1. The paper provided the first sample and computational complexity on testably learning PTFs, showing that testably learning PTFs is qualitatively similar in hardness to agnostic learning PTFs. To reach this result, the authors overcome a handful of technical obstacles that arose in adapting the fooling techniques to testable learning. Critically, the authors constructed a new low-degree polynomial to approximate the PTFs based on [Kane11] with refined approximation error bounds. The techniques the authors applied here could be of independent interest.

Weaknesses

1. Though the final sample complexity is $n^{poly(1/\epsilon)}$, the $poly(1/\epsilon)$ is of order $\epsilon^{-7^d}$, in other words, it is substantially worse than agnostic learning in terms of the order of $1/\epsilon$. It might be too harsh to say this is a serious weakness of this paper, as this is the first paper that provided these kinds of complex results; I think it would be an interesting future work to reduce the order of $1/\epsilon$ that is truly comparable to agnostic learning.

Questions

see weakness above.

Rating

6

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

the authors have addressed the limitations properly.

Reviewer 2v3u5/10 · confidence 4/52024-07-13

Summary

The authors study the problem of testing polynomial threshold functions in the agnostic setting in the testable learning paradigm. They present a testable learning algorithm that matches with the asymptotic bound (in terms of n) known for agnostic learning

Strengths

Testing of polynomial threshold functions is a very natural problem and the testable learning paradigm is also very natural. The authors present the first such testable learning algorithm in the agnostic learning setting for PTFs.

Weaknesses

The dependence on epsilon and the degree is very bad. This makes the results completely useless is practice. In fact the degree os the polynomial on the exponent is exponentially dependent on d.

Questions

Can you please argue why the dependence on d is so bad and is there possibility of improving it. Can a better bounds be obtained for smaller d, like d=2,3.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

NA

Reviewer pSUk6/10 · confidence 3/52024-07-17

Summary

This paper investigates testable learning of polynomial threshold functions (PTFs) with respect to the standard Gaussian distribution. The authors extend previous work on testable learning of halfspaces to show that PTFs of arbitrary constant degree can be testably learned up to excess error \epsilon in time n^poly(1/\epsilon), matching the best known guarantees in the agnostic model. The key technical contribution is showing that distributions which approximately match the moments of a Gaussian up to degree poly(1/\epsilon) fool constant-degree PTFs.

Strengths

- The paper is well written and easy to follow. - The paper makes significant progress on an open problem in learning theory by extending testable learning to PTFs.

Weaknesses

- It is not clear how important the problem of testable learning for PTFs is and if the results and/or techniques have applicability to other learning theory problems.

Questions

- It seems like the testable learning setting is similar to the samples coming from a distribution close to the distribution D in some sense? Is there any model of learning theory that explicitly studies this? - Do you see a path to improving the sample complexity to polynomial in both n and 1/\epsilon, rather than n^poly(1/\epsilon)? - The authors have looked at Gaussian distribution in this work. What other distributions could this result be extended to? - Can you provide intuition for why the runtime dependence on d is so much worse than in the agnostic model? Do you believe this gap is inherent or an artifact of the analysis? - Did the authors look into getting any sort of lower bounds for this problem?

Rating

6

Confidence

3

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes

Reviewer 9MWZ2024-08-10

Thank you for your response. I think this paper will be worth reading for people interested in testable learning and more enjoyable if the authors improving their writing in the final version in the way they claimed in this rebuttal. I raised my score to 7.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC