We study the problem of designing hard negative sampling distributions for unsupervised contrastive representation learning. We propose and analyze a novel min-max framework that seeks a representation which minimizes the maximum (worst-case) generalized contrastive learning loss over all couplings (joint distributions between positive and negative samples subject to marginal constraints) and prove that the resulting min-max optimum representation will be degenerate. This provides the first theoretical justification for incorporating additional regularization constraints on the couplings. We re-interpret the min-max problem through the lens of Optimal Transport (OT) theory and utilize regularized transport couplings to control the degree of hardness of negative examples. Through experiments we demonstrate that the negative samples generated from our designed negative distribution are more similar to the anchor than those generated from the baseline negative distribution. We also demonstrate that entropic regularization yields negative sampling distributions with parametric form similar to that in a recent state-of-the-art negative sampling design and has similar performance in multiple datasets. Utilizing the uncovered connection with OT, we propose a new ground cost for designing the negative distribution and show improved performance of the learned representation on downstream tasks compared to the representation learned when using squared Euclidean cost.

For a widely-studied data model and general loss and sample-hardening functions we prove that the Supervised Contrastive Learning (SCL), Hard-SCL (HSCL), and Unsupervised Contrastive Learning (UCL) risks are minimized by representations that exhibit Neural Collapse (NC), i.e., the class means form an Equianglular Tight Frame (ETF) and data from the same class are mapped to the same representation. We also prove that for any representation mapping, the HSCL and Hard-UCL (HUCL) risks are lower bounded by the corresponding SCL and UCL risks. Although the optimality of ETF is known for SCL, albeit only for InfoNCE loss, its optimality for HSCL and UCL under general loss and hardening functions is novel. Moreover, our proofs are much simpler, compact, and transparent. We empirically demonstrate, for the first time, that ADAM optimization of HSCL and HUCL risks with random initialization and suitable hardness levels can indeed converge to the NC geometry if we incorporate unit-ball or unit-sphere feature normalization. Without incorporating hard negatives or feature normalization, however, the representations learned via ADAM suffer from dimensional collapse (DC) and fail to attain the NC geometry.

Domain generalization aims to learn a model with good generalization ability, that is, the learned model should not only perform well on several seen domains but also on unseen domains with different data distributions. State-of-the-art domain generalization methods typically train a representation function followed by a classifier jointly to minimize both the classification risk and the domain discrepancy. However, when it comes to model selection, most of these methods rely on traditional validation routines that select models solely based on the lowest classification risk on the validation set. In this paper, we theoretically demonstrate a trade-off between minimizing classification risk and mitigating domain discrepancy, i.e., it is impossible to achieve the minimum of these two objectives simultaneously. Motivated by this theoretical result, we propose a novel model selection method suggesting that the validation process should account for both the classification risk and the domain discrepancy. We validate the effectiveness of the proposed method by numerical results on several domain generalization datasets.

Invariance principle-based methods, for example, Invariant Risk Minimization (IRM), have recently emerged as promising approaches for Domain Generalization (DG). Despite the promising theory, invariance principle-based approaches fail in common classification tasks due to the mixture of the true invariant features and the spurious invariant features. In this paper, we propose a framework based on the conditional entropy minimization principle to filter out the spurious invariant features leading to a new algorithm with a better generalization capability. We theoretically prove that under some particular assumptions, the representation function can precisely recover the true invariant features. In addition, we also show that the proposed approach is closely related to the well-known Information Bottleneck framework. Both the theoretical and numerical results are provided to justify our approach.

Random walk based node embedding algorithms learn vector representations of nodes by optimizing an objective function of node embedding vectors and skip-bigram statistics computed from random walks on the network. They have been applied to many supervised learning problems such as link prediction and node classification and have demonstrated state-of-the-art performance. Yet, their properties remain poorly understood. This paper studies properties of random walk based node embeddings in the unsupervised setting of discovering hidden block structure in the network, i.e., learning node representations whose cluster structure in Euclidean space reflects their adjacency structure within the network. We characterize the ergodic limits of the embedding objective, its generalization, and related convex relaxations to derive corresponding non-randomized versions of the node embedding objectives. We also characterize the optimal node embedding Grammians of the non-randomized objectives for the expected graph of a two-community Stochastic Block Model (SBM). We prove that the solution Grammian has rank $1$ for a suitable nuclear norm relaxation of the non-randomized objective. Comprehensive experimental results on SBM random networks reveal that our non-randomized ergodic objectives yield node embeddings whose distribution is Gaussian-like, centered at the node embeddings of the expected network within each community, and concentrate in the linear degree-scaling regime as the number of nodes increases.

For the Domain Generalization (DG) problem where the hypotheses are composed of a common representation function followed by a labeling function, we point out a shortcoming in existing approaches that fail to explicitly optimize for a term, appearing in a well-known and widely adopted upper bound to the risk on the unseen domain, that is dependent on the representation to be learned. To this end, we first derive a novel upper bound to the prediction risk. We show that imposing a mild assumption on the representation to be learned, namely manifold restricted invertibility, is sufficient to deal with this issue. Further, unlike existing approaches, our novel upper bound doesn't require the assumption of Lipschitzness of the loss function. In addition, the distributional discrepancy in the representation space is handled via the Wasserstein-2 barycenter cost. In this context, we creatively leverage old and recent transport inequalities, which link various optimal transport metrics, in particular the $L^1$ distance (also known as the total variation distance) and the Wasserstein-2 distances, with the Kullback-Liebler divergence. These analyses and insights motivate a new representation learning cost for DG that additively balances three competing objectives: 1) minimizing classification error across seen domains via cross-entropy, 2) enforcing domain-invariance in the representation space via the Wasserstein-2 barycenter cost, and 3) promoting non-degenerate, nearly-invertible representation via one of two mechanisms, viz., an autoencoder-based reconstruction loss or a mutual information loss. It is to be noted that the proposed algorithms completely bypass the use of any adversarial training mechanism that is typical of many current domain generalization approaches. Simulation results on several standard datasets demonstrate superior performance compared to several well-known DG algorithms.

Due to concerns about human error in crowdsourcing, it is standard practice to collect labels for the same data point from multiple internet workers. We here show that the resulting budget can be used more effectively with a flexible worker assignment strategy that asks fewer workers to analyze easy-to-label data and more workers to analyze data that requires extra scrutiny. Our main contribution is to show how the allocations of the number of workers to a task can be computed optimally based on task features alone, without using worker profiles. Our target tasks are delineating cells in microscopy images and analyzing the sentiment toward the 2016 U.S. presidential candidates in tweets. We first propose an algorithm that computes budget-optimized crowd worker allocation (BUOCA). We next train a machine learning system (BUOCA-ML) that predicts an optimal number of crowd workers needed to maximize the accuracy of the labeling. We show that the computed allocation can yield large savings in the crowdsourcing budget (up to 49 percent points) while maintaining labeling accuracy. Finally, we envisage a human-machine system for performing budget-optimized data analysis at a scale beyond the feasibility of crowdsourcing.

We propose a data-driven framework for optimizing privacy-preserving data release mechanisms toward the information-theoretically optimal tradeoff between minimizing distortion of useful data and concealing sensitive information. Our approach employs adversarially-trained neural networks to implement randomized mechanisms and to perform a variational approximation of mutual information privacy. We empirically validate our Privacy-Preserving Adversarial Networks (PPAN) framework with experiments conducted on discrete and continuous synthetic data, as well as the MNIST handwritten digits dataset. With the synthetic data, we find that our model-agnostic PPAN approach achieves tradeoff points very close to the optimal tradeoffs that are analytically-derived from model knowledge. In experiments with the MNIST data, we visually demonstrate a learned tradeoff between minimizing the pixel-level distortion versus concealing the written digit.

Neural node embeddings have recently emerged as a powerful representation for supervised learning tasks involving graph-structured data. We leverage this recent advance to develop a novel algorithm for unsupervised community discovery in graphs. Through extensive experimental studies on simulated and real-world data, we demonstrate that the proposed approach consistently improves over the current state-of-the-art. Specifically, our approach empirically attains the information-theoretic limits for community recovery under the benchmark Stochastic Block Models for graph generation and exhibits better stability and accuracy over both Spectral Clustering and Acyclic Belief Propagation in the community recovery limits.

We develop necessary and sufficient conditions and a novel provably consistent and efficient algorithm for discovering topics (latent factors) from observations (documents) that are realized from a probabilistic mixture of shared latent factors that have certain properties. Our focus is on the class of topic models in which each shared latent factor contains a novel word that is unique to that factor, a property that has come to be known as separability. Our algorithm is based on the key insight that the novel words correspond to the extreme points of the convex hull formed by the row-vectors of a suitably normalized word co-occurrence matrix. We leverage this geometric insight to establish polynomial computation and sample complexity bounds based on a few isotropic random projections of the rows of the normalized word co-occurrence matrix. Our proposed random-projections-based algorithm is naturally amenable to an efficient distributed implementation and is attractive for modern web-scale distributed data mining applications.