Agnostic Learning of Arbitrary ReLU Activation under Gaussian Marginals

We consider the problem of learning an arbitrarily-biased ReLU activation (or neuron) over Gaussian marginals with the squared loss objective. Despite the ReLU neuron being the basic building block of modern neural networks, we still do not understand the basic algorithmic question of whether one arbitrary ReLU neuron is learnable in the non-realizable setting. In particular, all existing polynomial time algorithms only provide approximation guarantees for the better-behaved unbiased setting or restricted bias setting. Our main result is a polynomial time statistical query (SQ) algorithm that gives the first constant factor approximation for arbitrary bias. It outputs a ReLU activation that achieves a loss of $O(\mathrm{OPT}) + \varepsilon$ in time $\mathrm{poly}(d,1/\varepsilon)$, where $\mathrm{OPT}$ is the loss obtained by the optimal ReLU activation. Our algorithm presents an interesting departure from existing algorithms, which are all based on gradient descent and thus fall within the class of correlational statistical query (CSQ) algorithms. We complement our algorithmic result by showing that no polynomial time CSQ algorithm can achieve a constant factor approximation. Together, these results shed light on the intrinsic limitation of gradient descent, while identifying arguably the simplest setting (a single neuron) where there is a separation between SQ and CSQ algorithms.

Theoretical Analysis of Weak-to-Strong Generalization

Strong student models can learn from weaker teachers: when trained on the predictions of a weaker model, a strong pretrained student can learn to correct the weak model's errors and generalize to examples where the teacher is not confident, even when these examples are excluded from training. This enables learning from cheap, incomplete, and possibly incorrect label information, such as coarse logical rules or the generations of a language model. We show that existing weak supervision theory fails to account for both of these effects, which we call pseudolabel correction and coverage expansion, respectively. We give a new bound based on expansion properties of the data distribution and student hypothesis class that directly accounts for pseudolabel correction and coverage expansion. Our bounds capture the intuition that weak-to-strong generalization occurs when the strong model is unable to fit the mistakes of the weak teacher without incurring additional error. We show that these expansion properties can be checked from finite data and give empirical evidence that they hold in practice.

Efficient Certificates of Anti-Concentration Beyond Gaussians

A set of high dimensional points $X=\{x_1, x_2,\ldots, x_n\} \subset R^d$ in isotropic position is said to be $\delta$-anti concentrated if for every direction $v$, the fraction of points in $X$ satisfying $|\langle x_i,v \rangle |\leq \delta$ is at most $O(\delta)$. Motivated by applications to list-decodable learning and clustering, recent works have considered the problem of constructing efficient certificates of anti-concentration in the average case, when the set of points $X$ corresponds to samples from a Gaussian distribution. Their certificates played a crucial role in several subsequent works in algorithmic robust statistics on list-decodable learning and settling the robust learnability of arbitrary Gaussian mixtures, yet remain limited to rotationally invariant distributions. This work presents a new (and arguably the most natural) formulation for anti-concentration. Using this formulation, we give quasi-polynomial time verifiable sum-of-squares certificates of anti-concentration that hold for a wide class of non-Gaussian distributions including anti-concentrated bounded product distributions and uniform distributions over $L_p$ balls (and their affine transformations). Consequently, our method upgrades and extends results in algorithmic robust statistics e.g., list-decodable learning and clustering, to such distributions. Our approach constructs a canonical integer program for anti-concentration and analysis a sum-of-squares relaxation of it, independent of the intended application. We rely on duality and analyze a pseudo-expectation on large subsets of the input points that take a small value in some direction. Our analysis uses the method of polynomial reweightings to reduce the problem to analyzing only analytically dense or sparse directions.

Error-Tolerant E-Discovery Protocols

We consider the multi-party classification problem introduced by Dong, Hartline, and Vijayaraghavan (2022) in the context of electronic discovery (e-discovery). Based on a request for production from the requesting party, the responding party is required to provide documents that are responsive to the request except for those that are legally privileged. Our goal is to find a protocol that verifies that the responding party sends almost all responsive documents while minimizing the disclosure of non-responsive documents. We provide protocols in the challenging non-realizable setting, where the instance may not be perfectly separated by a linear classifier. We demonstrate empirically that our protocol successfully manages to find almost all relevant documents, while incurring only a small disclosure of non-responsive documents. We complement this with a theoretical analysis of our protocol in the single-dimensional setting, and other experiments on simulated data which suggest that the non-responsive disclosure incurred by our protocol may be unavoidable.

Computing linear sections of varieties: quantum entanglement, tensor decompositions and beyond

We study the problem of finding elements in the intersection of an arbitrary conic variety in $\mathbb{F}^n$ with a given linear subspace (where $\mathbb{F}$ can be the real or complex field). This problem captures a rich family of algorithmic problems under different choices of the variety. The special case of the variety consisting of rank-1 matrices already has strong connections to central problems in different areas like quantum information theory and tensor decompositions. This problem is known to be NP-hard in the worst-case, even for the variety of rank-1 matrices. Surprisingly, despite these hardness results we give efficient algorithms that solve this problem for "typical" subspaces. Here, the subspace $U \subseteq \mathbb{F}^n$ is chosen generically of a certain dimension, potentially with some generic elements of the variety contained in it. Our main algorithmic result is a polynomial time algorithm that recovers all the elements of $U$ that lie in the variety, under some mild non-degeneracy assumptions on the variety. As corollaries, we obtain the following results: $\bullet$ Uniqueness results and polynomial time algorithms for generic instances of a broad class of low-rank decomposition problems that go beyond tensor decompositions. Here, we recover a decomposition of the form $\sum_{i=1}^R v_i \otimes w_i$, where the $v_i$ are elements of the given variety $X$. This implies new algorithmic results even in the special case of tensor decompositions. $\bullet$ Polynomial time algorithms for several entangled subspaces problems in quantum entanglement, including determining $r$-entanglement, complete entanglement, and genuine entanglement of a subspace. While all of these problems are NP-hard in the worst case, our algorithm solves them in polynomial time for generic subspaces of dimension up to a constant multiple of the maximum possible.

Classification Protocols with Minimal Disclosure

We consider multi-party protocols for classification that are motivated by applications such as e-discovery in court proceedings. We identify a protocol that guarantees that the requesting party receives all responsive documents and the sending party discloses the minimal amount of non-responsive documents necessary to prove that all responsive documents have been received. This protocol can be embedded in a machine learning framework that enables automated labeling of points and the resulting multi-party protocol is equivalent to the standard one-party classification problem (if the one-party classification problem satisfies a natural independence-of-irrelevant-alternatives property). Our formal guarantees focus on the case where there is a linear classifier that correctly partitions the documents.

Agnostic Learning of General ReLU Activation Using Gradient Descent

We provide a convergence analysis of gradient descent for the problem of agnostically learning a single ReLU function under Gaussian distributions. Unlike prior work that studies the setting of zero bias, we consider the more challenging scenario when the bias of the ReLU function is non-zero. Our main result establishes that starting from random initialization, in a polynomial number of iterations gradient descent outputs, with high probability, a ReLU function that achieves a competitive error guarantee when compared to the error of the best ReLU function. We also provide finite sample guarantees, and these techniques generalize to a broader class of marginal distributions beyond Gaussians.

Training Subset Selection for Weak Supervision

Existing weak supervision approaches use all the data covered by weak signals to train a classifier. We show both theoretically and empirically that this is not always optimal. Intuitively, there is a tradeoff between the amount of weakly-labeled data and the precision of the weak labels. We explore this tradeoff by combining pretrained data representations with the cut statistic (Muhlenbach et al., 2004) to select (hopefully) high-quality subsets of the weakly-labeled training data. Subset selection applies to any label model and classifier and is very simple to plug in to existing weak supervision pipelines, requiring just a few lines of code. We show our subset selection method improves the performance of weak supervision for a wide range of label models, classifiers, and datasets. Using less weakly-labeled data improves the accuracy of weak supervision pipelines by up to 19% (absolute) on benchmark tasks.

Efficient Algorithms for Learning Depth-2 Neural Networks with General ReLU Activations

We present polynomial time and sample efficient algorithms for learning an unknown depth-2 feedforward neural network with general ReLU activations, under mild non-degeneracy assumptions. In particular, we consider learning an unknown network of the form $f(x) = {a}^{\mathsf{T}}\sigma({W}^\mathsf{T}x+b)$, where $x$ is drawn from the Gaussian distribution, and $\sigma(t) := \max(t,0)$ is the ReLU activation. Prior works for learning networks with ReLU activations assume that the bias $b$ is zero. In order to deal with the presence of the bias terms, our proposed algorithm consists of robustly decomposing multiple higher order tensors arising from the Hermite expansion of the function $f(x)$. Using these ideas we also establish identifiability of the network parameters under minimal assumptions.

Beyond Perturbation Stability: LP Recovery Guarantees for MAP Inference on Noisy Stable Instances

Several works have shown that perturbation stable instances of the MAP inference problem in Potts models can be solved exactly using a natural linear programming (LP) relaxation. However, most of these works give few (or no) guarantees for the LP solutions on instances that do not satisfy the relatively strict perturbation stability definitions. In this work, we go beyond these stability results by showing that the LP approximately recovers the MAP solution of a stable instance even after the instance is corrupted by noise. This "noisy stable" model realistically fits with practical MAP inference problems: we design an algorithm for finding "close" stable instances, and show that several real-world instances from computer vision have nearby instances that are perturbation stable. These results suggest a new theoretical explanation for the excellent performance of this LP relaxation in practice.