A Near-optimal Algorithm for Learning Margin Halfspaces with Massart Noise

We study the problem of PAC learning $γ$-margin halfspaces in the presence of Massart noise. Without computational considerations, the sample complexity of this learning problem is known to be $\widetildeΘ(1/(γ^2 ε))$. Prior computationally efficient algorithms for the problem incur sample complexity $\tilde{O}(1/(γ^4 ε^3))$ and achieve 0-1 error of $η+ε$, where $η<1/2$ is the upper bound on the noise rate. Recent work gave evidence of an information-computation tradeoff, suggesting that a quadratic dependence on $1/ε$ is required for computationally efficient algorithms. Our main result is a computationally efficient learner with sample complexity $\widetildeΘ(1/(γ^2 ε^2))$, nearly matching this lower bound. In addition, our algorithm is simple and practical, relying on online SGD on a carefully selected sequence of convex losses.

Paper

References (37)

Scroll for more · 25 remaining

Similar papers

Peer review

Reviewer 2xNf7/10 · confidence 3/52024-06-27

Summary

The paper considers the problem of PAC learning halfspaces with margin in the presence of Massart noise. The paper provides an algorithm that well-balances sample and computational efficiency. Specifically, the dependence of the algorithm on both $\epsilon$ and $\gamma$ is near-optimal.

Strengths

1. The studied problem is fundamental in the area of PAC learning, and the paper provides a significant progress on it. 2. The paper is well written, and the main techniques are well explained.

Weaknesses

1. Results might be somewhat weaker than presented (see questions), and some phrasings in this context are too vague (for example "there is evidence that..."). 2. The natural agnostic extension of the problem is not discussed.

Questions

Questions: 1. I'm not sure about the statement "...we essentially settle this question..."  in line 71. As far as I understand, the optimal computational sample complexity might be $\tilde{O}(1/\epsilon^2 + 1/\epsilon \gamma^2)$ and not $\tilde{\Omega}(1/\epsilon^2 \gamma^2)$. Either way, the provided upper bound is impressive enough to recommend for acceptance. 2. Is there a specific reason not to discuss the agnostic case? Is it usualy considered in Massart noise problems? Suggestions: Consider writing what is $\eta$ in the abstract.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes.

Reviewer Sz9S7/10 · confidence 4/52024-07-05

Summary

The paper considers the problem of PAC-learning $\gamma$-margin halfspaces under $\eta$-Massart noise. The paper provides an efficient algorithm achieving error $\eta+\epsilon$ with sample complexity $\tilde{O}(1/(\epsilon^2\gamma^2))$. Since previous work provided evidence that an inverse-quadratic dependence on $\epsilon$ is necessary for efficient algorithms, the algorithm appears to be near-optimal in terms of sample-complexity up to logarithmic factors. The algorithm relies on an iterative stochastic-gradient descent approach, where at each iteration a new loss function is defined which determines the gradient descent of the next iteration. The paper proves that if $T$ iterations are performed (where $T$ is a sufficiently large multiple of $\log(1/\delta)/(\epsilon^2\gamma^2)$), then with probability at least $1-\delta$, one of the halfspaces obtained in $T$ iterations have error at most $\eta+\epsilon$. By drawing some extra independent samples and computing the empirical error for each of these halfspaces, one can find a good halfspace. The paper is generally well written.

Strengths

The paper provides a near-optimal algorithm (in terms of sample complexity) for learning $\gamma$-margin halfspaces under Massart noise.

Weaknesses

I haven't found significant weaknesses in the paper. Typos: - Page 2, line 34: “We say that that the distribution” -> “We say that the distribution” - Page 5: Line 195: If I am not mistaken, the function g is twice the gradient.

Questions

- The authors claim that the computational complexity of the algorithm is linear in the samples. However, it seems to me that the complexity of step (5) is $O(T\cdot N\cdot d)$. It does not seem to be trivial to implement step (5) in $O((T + N)\cdot d)$ time. - While previous work show that $\Omega(1/(\gamma^2\epsilon))$ and $\Omega(1/\epsilon^2)$ are lower bounds, it doesn't necessarily follow that $\Omega(1/(\epsilon^2\gamma^2))$ is a lower bound as well.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Since this is a theoretical paper, I can't see any potential negative societal impact.

Reviewer JiXV7/10 · confidence 3/52024-07-12

Summary

This paper studies the problem of PAC learning $\gamma$-margin halfspaces under Massart noise: to PAC learn any distribution $D$ such that there exists $w^*$ with the bounded norm of $1$ that has a margin of at least $\gamma$, i.e. $\mathbb{P}_{(x,y)\sim D}( |\langle w^*,x \rangle| \geq \gamma)=1$ and for every $x$, we have $\mathbb{P}(\text{sign}(\langle w^*,x \rangle) \neq y )=\eta(x)$, where a function $\eta$ is bounded in $[0,\eta]$ (with overload of notation). This is a classical problem in learning theory. Information theoretically, one needs $\tilde{\Theta}(1/\gamma^2 \epsilon)$ samples, while for computationally efficient algorithms, it is widely believed that inverse quadratic dependence in $\epsilon$ (e.g. $1/\epsilon^2$) is necessary. The main result of this paper is to present an efficient algorithm with nearly optimal sample complexity $\tilde{O}(1/\gamma^2 \epsilon^2)$, essentially closing the question.

Strengths

1. The paper studies a problem that is important to the learning theory community and provides a state-of-the-art result. The previous best algorithm required $O(1/\gamma^4 \epsilon^3)$ samples. 2. The paper is well-written.

Weaknesses

I do not see any major weaknesses.

Questions

None.

Rating

7

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

This is a theoretical work with no societal impact.

Reviewer SyXT7/10 · confidence 3/52024-07-13

Summary

The paper studies the problem of learning a $\\gamma$-margin halfspace with $\\eta$ Massart noise and provides the first computationally efficient algorithm having 0-1 error $<= \\eta + \\varepsilon$ with sample complexity $O(1/\\gamma^2\\varepsilon^2)$ which nearly matches the information theoretic bound of $O(1/\\gamma^2\\varepsilon^)$ improving upon the previous $O(1/\\gamma^4\\varepsilon^3)$ sample complexity by [Chen et al. ‘20]. The main contributions of the paper are a novel “global” (as opposed to previous conditioning based) optimization to solve this problem directly using gradient descent. Towards this, a novel convex loss is defined using an independent vector as a parameter, and using the margin as a bound on the scaling parameter. The gradient w.r.t. the solution vector is used in an SGD loop, while the independent vector is updated by the gradient step, and its projection on the unit ball is the updated solution. This clever formulation leads to the gradient being composed of the “correct” direction of optimization along with an estimation error. The former minimizes the distance to the true solution, while the latter is bounded using standard concentration arguments.

Strengths

1. Novel loss formulation and gradient update step which can be directly used via SGD. 2. Simple algorithm and analysis. 3. Near optimal computational bound on an important problem.

Weaknesses

Result is specific to a particular noise model and it is not clear if the techniques are more broadly applicable.

Questions

In Algorithm 1, $\\lambda_t$ seems to be a constant independent of $t$, so it can be fixed outside the loop.

Rating

7

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

Yes

Reviewer g5LR8/10 · confidence 3/52024-07-15

Summary

This paper essentially resolves the *computational sample complexity* of learning $\gamma$-margin halfspaces under the Massart noise model. Here, computational sample complexity refers to the number of samples required for polynomial-time algorithms, as opposed to general statistical estimators which may be computationally intractable. Previous works have given rigorous evidence, in the form of lower bounds against large classes of algorithms such as SQ algorithms, of an *information-computation* gap for learning $\gamma$-margin halfspaces. Without restrictions on computation, the sample complexity is $\Theta(1/(\gamma^2 \epsilon))$ for achieving zero-one loss of $\eta + \epsilon$, where $\eta$ is the uniform bound on the Massart noise. For a large class of efficient algorithms, however, the required sample complexity is at least $\Omega(1/(\gamma^2 \epsilon^2))$. They show that the $1/(\gamma^2 \epsilon^2)$ sample complexity is essentially tight by designing a polynomial-time algorithm for learning $\gamma$-margin halfspaces. Their algorithm is based on a novel and technically insightful choice of the loss sequence for online SGD, which results in a surprisingly simple and efficient algorithm with near-optimal sample complexity.

Strengths

This is a well-written paper that stands out for being both technically insightful and simple. A simple and efficient algorithm (online SGD with a specific choice of loss sequence) achieves optimal sample complexity and lends itself to a clean analysis. It's hard to ask for more than this. The core idea behind the algorithm is a clever choice of the loss sequence $(\ell_t)$ with respect to which one runs online SGD. Previous works have already employed the LeakyReLU loss as a convex surrogate for the zero-one loss and achieved suboptimal upper bounds on the sample complexity. LeakyReLU with parameter $\lambda > 0$ is defined by $\ell_\lambda(a) = (1-\lambda) \mathbb{1}[a \ge 0] + \lambda \mathbb{1}[a < 0]$. Applying this to the halfspace learning setting, a straightforward calculation shows that $\ell_\lambda(-y (w \cdot x)) = (\mathbb{1}[\mathrm{sign}(w\cdot x) \neq y] - \lambda)|w \cdot x|$, where $(x, y) \in \mathbb{S}^{d-1} \times \\{\pm 1\\}$ is a sample from the $\gamma$-margin halfspace distribution and $w \in \mathbb{R}^d$ is a candidate halfspace. Note the resemblance to the *shifted* zero-one loss $L(w) = \mathbb{E}_{(x,y)} \mathbb{1}[\mathrm{sign}(w\cdot x) \neq y] - \eta$ (when $\lambda = \eta$ in LeakyReLU). By the $\eta$-Massart noise assumption, $L(w^{\*}) \le 0$ for the optimal halfspace $w^{\*} \in \mathbb{R}^d$ and $L(w) \ge \epsilon$ for any halfspace $w$ with zero-one loss at least $\eta + \epsilon$. The key difference between this shifted zero-one loss and the LeakyReLU is the $|w \cdot x|$ term. In particular, if we reweight each sample LeakyReLU loss by $1/|w \cdot x|$, we recover the shifted zero-one loss which is unfortunately non-convex. The authors overcome this issue by considering a *family* of bounded and convex loss functions $(\ell_u)$ indexed by $u \in \mathbb{R}^{d}$. Each $\ell_u$ is simply the LeakyReLU loss reweighted by $1/|u \cdot x|$. It is precisely this decoupling of the halfspace parameter $w$ and the reweighting parameter $u$ that leads to the guarantees of the algorithm. The *sequence* of reweighting parameters $(u_t)$, each of which leads to a different loss $\ell_t$, is chosen adaptively by online SGD.

Weaknesses

I did not find any significant weaknesses, only minor comments regarding the presentation. - **The expression "(vector) v is independent of w" (line 174, 176) is confusing.** I think it's easy to confuse "independence" with statistical independence. It would be helpful to clarify this by adding that the reweighting term is *constant* with respect to the parameter $w \in \mathbb{R}^d$, and the reweighting term $W(v \cdot x, \gamma)$ remains the same when taking the gradient of $L_{v}(w)$. This is already implicit in the mathematical expressions, but providing additional explanation would benefit readers.

Questions

- Does the near-optimal online SGD algorithm fall within the class of SQ algorithms? If not, could there still exist non-SQ algorithms that achieve sample complexity with subquadratic dependence on $1/\epsilon$?

Rating

8

Confidence

3

Soundness

4

Presentation

4

Contribution

3

Limitations

N/A

Reviewer Sz9S2024-08-11

I would like to thank the authors for their response. Since they addressed my minor comments, I'm increasing my score to 7.

Reviewer SyXT2024-08-11

Thank you for the rebuttal and the clarification. I will keep my rating.

Program Chairsdecision2024-09-25

Decision

Accept (spotlight)

© 2026 NYSGPT2525 LLC