The study develops a theory for structured prediction using smooth convex surrogates.
problem Developing a theoretical framework for structured prediction.
method Characterizing smooth convex surrogates compatible with task losses and deriving statistical guarantees.
result Derives tight bounds for the calibration function and novel results for existing surrogate frameworks.
We establish linear regret bounds for convex smooth losses using Fenchel-Young losses.
problem Establishing linear regret bounds for convex smooth losses.
method Constructing a convex smooth surrogate loss using Fenchel-Young losses generated by the convolutional negentropy.
result We derive a smooth loss with a linear surrogate regret bound.
In statistical learning theory, convex surrogates of the 0-1 loss are highly preferred because of the computational and theoretical virtues that convexity brings in. This is of more importance if we consider smooth surrogates as witnessed by the fact that the smoothness is further beneficial both computationally- by at…
Optimizes hard-to-optimize metrics using adaptive surrogates.
problem Training models with black-box and hard-to-optimize metrics.
method Expresses metric as a function of surrogates, solves optimization problem over relaxed surrogate space.
result Approach performs on par with known methods and adds value when metric form is unknown.
Signal estimation problems with smoothness and sparsity priors can be naturally modeled as quadratic optimization with ℓ0-"norm" constraints. Since such problems are non-convex and hard-to-solve, the standard approach is, instead, to tackle their convex surrogates based on ℓ1-norm relaxations. In this paper…
Linear-Core Surrogates combine fast optimization and statistical efficiency in classification and structured prediction.
problem The trade-off between smoothness and margin-based losses in classification and structured prediction.
method Linear-Core (LC) Surrogates, a family of convex loss functions that stitch a linear core to a smooth tail.
result LC Surrogates achieve fast linear consistency rates while maintaining differentiability and strict H-consistency bounds. New algorithm learns optimal stepsizes for SGD in noisy non-convex optimization.
problem Finding optimal stepsize for SGD in noisy non-convex optimization.
method Surrogate losses cast problem into online convex optimization, using no-regret algorithms.
result Self-tuned SGD algorithm with adaptive convergence rates.
New method smooths optimization for sparse regularization.
problem Non-smooth, non-convex optimization problems for sparsity.
method Overparameterization and smooth surrogate penalties.
result Surrogate objective has identical global and local minima.
We carefully study how well minimizing convex surrogate loss functions, corresponds to minimizing the misclassification error rate for the problem of binary classification with linear predictors. In particular, we show that amongst all convex surrogate losses, the hinge loss gives essentially the best possible bound, o…
Empirical risk minimization frequently employs convex surrogates to underlying discrete loss functions in order to achieve computational tractability during optimization. However, classical convex surrogates can only tightly bound modular loss functions, sub-modular functions or supermodular functions separately while …
New algorithms reduce regret for convex bandits with small comparator norms.
problem Optimizing in bandit convex optimization with varying comparator norms.
method Developed algorithms using techniques from full-information setting and new gradient estimators.
result Regret bounds are small when comparator norm is small.
Learning with non-modular losses is an important problem when sets of predictions are made simultaneously. The main tools for constructing convex surrogate loss functions for set prediction are margin rescaling and slack rescaling. In this work, we show that these strategies lead to tight convex surrogates iff the unde…
Study on H-consistency bounds for machine learning surrogates.
problem Estimating target loss error relative to surrogate loss error in machine learning.
method Developed H-consistency bounds for various surrogates and loss functions. result Stronger guarantees than existing methods, offering distribution-dependent and -independent bounds.
Paper extends SMM to weakly convex and multi-convex surrogates for non-convex optimization.
problem Non-convex optimization with weakly convex or multi-convex surrogates.
method Stochastic majorization-minimization with proximal regularization or block-minimization.
result Convergence rates for empirical and expected losses under non-i.i.d. data.
Finding efficient and provable methods to solve non-convex optimization problems is an outstanding challenge in machine learning and optimization theory. A popular approach used to tackle non-convex problems is to use convex relaxation techniques to find a convex surrogate for the problem. Unfortunately, convex relaxat…
We learn a compact surrogate model for optimization problems to reduce training and inference time.
problem Solving optimization problems with unknown parameters is computationally expensive and may lead to suboptimal solutions.
method We represent the optimization problem in terms of meta-variables and learn a low-dimensional surrogate model end-to-end with the predictive model.
result We achieve a large reduction in training and inference time, and improved performance.
We study consistency properties of surrogate loss functions for general multiclass learning problems, defined by a general multiclass loss matrix. We extend the notion of classification calibration, which has been studied for binary and multiclass 0-1 classification problems (and for certain other specific learning pro…
This research analyzes the consistency of convex and nonconvex surrogate losses for adversarially robust classification.
problem Ensuring classifiers are robust to adversarial perturbations.
method Analysis of convex and nonconvex surrogate losses through the lens of calibration.
result No convex surrogate loss is calibrated with respect to the adversarial 0-1 loss for linear models, but nonconvex losses can be calibrated under certain conditions.
New model guarantees integer optimal solutions for combinatorial problems.
problem Finding optimal solutions for combinatorial problems with costly or noisy evaluations.
method Developed a surrogate model with integer-valued minima for combinatorial optimization.
result Outperforms other optimization algorithms on specific combinatorial problems.
Adversarial consistency depends on the uniqueness of adversarial Bayes classifiers.
problem Consistency of adversarial surrogate losses is not guaranteed.
method Connected consistency of adversarial surrogate losses to the uniqueness of adversarial Bayes classifiers.
result A convex surrogate loss is statistically consistent for adversarial learning if and only if the adversarial Bayes classifier is unique.
We provide novel theoretical insights on structured prediction in the context of efficient convex surrogate loss minimization with consistency guarantees. For any task loss, we construct a convex surrogate that can be optimized via stochastic gradient descent and we prove tight bounds on the so-called "calibration func…
We propose a robust adversarial prediction framework for general multiclass classification. Our method seeks predictive distributions that robustly optimize non-convex and non-continuous multiclass loss metrics against the worst-case conditional label distributions (the adversarial distributions) that (approximately) m…
The paper explores trading off consistency and dimensionality in convex surrogates for multiclass classification.
problem Designing consistent surrogate losses for multiclass classification with high-dimensional outcomes.
method Investigates embedding outcomes into convex polytopes and examining consistency under low-noise assumptions.
result Consistency can be achieved with less than n−1 dimensions, but hallucination occurs for some distributions. We study consistency properties of machine learning methods based on minimizing convex surrogates. We extend the recent framework of Osokin et al. (2017) for the quantitative analysis of consistency properties to the case of inconsistent surrogates. Our key technical contribution consists in a new lower bound on the ca…
Surrogate-based analysis of interactions via local effect smooths
problem Detecting and characterizing feature interactions in machine learning models
method Surrogate-based analysis using generalized additive models
result Empirical validation of effectiveness for pairwise interactions
This paper develops convex surrogates for optimizing the multi-label F-measure.
problem Optimizing the F-measure for multi-label classification is computationally hard.
method Designing convex surrogate losses calibrated for the F-measure.
result The F-measure for multi-label problems has a rank of at most s2+1. We develop a framework for consistent polyhedral surrogates in classification and prediction.
problem Designing consistent polyhedral surrogates for classification and prediction problems.
method Formalizing and studying embeddings of predictions as points in R^d, assigning original loss values, and convexifying to create surrogates.
result Established a strong connection between embeddings and polyhedral surrogates, providing constructions and proofs of consistency or inconsistency.
New method simplifies checking consistency of differentiable loss functions.
problem Verifying consistency of differentiable loss functions is difficult.
method Developed a new approach called strong indirect elicitation (strong IE) to simplify checking consistency.
result Strong IE is equivalent to calibration for strongly convex, differentiable surrogates.
Proposes a differentiable LSE-ICNN for modeling multi-well potentials.
problem Modeling multi-well potentials in various scientific domains.
method Log-sum-exponential (LSE) mixture of input convex neural network (ICNN) modes.
result Smooth surrogate that retains convexity within basins and allows gradient-based learning.
This paper proposes a mechanism to produce equivalent Lipschitz surrogates for zero-norm and rank optimization problems by means of the global exact penalty for their equivalent mathematical programs with an equilibrium constraint (MPECs). Specifically, we reformulate these combinatorial problems as equivalent MPECs by…
Paper develops algorithms for nonsmooth, nonconvex statistical learning problems.
problem Nonsmooth and nonconvex objectives in statistical learning.
method Bregman-surrogate algorithm framework, including local linear approximation, mirror descent, iterative thresholding, DC programming.
result Global convergence rates for nonconvex and nonsmooth objectives in high dimensions.
Paper establishes a universal growth rate for smooth surrogate losses in classification.
problem Analyzing growth rates of consistency bounds for various surrogate losses.
method Proves square-root growth rate for smooth margin-based losses; extends to multi-class classification.
result Demonstrates a universal square-root growth rate for smooth comp-sum and constrained losses.
Safe reinforcement learning with nonconvex constraints using convex approximations.
problem Safe reinforcement learning with nonlinear function approximation.
method Constructing surrogate convex constrained optimization problems by replacing nonconvex functions with convex quadratic functions.
result Solutions to surrogate problems converge to a stationary point of the original nonconvex problem.
New bounds show polyhedral surrogates are optimal for generalization.
problem Proving generalization rates for polyhedral loss functions.
method Developed two general results for polyhedral surrogates.
result Polyhedral surrogates provide linear surrogate regret bounds, translating directly to target rates.
We develop a variant of multiclass logistic regression that is significantly more robust to noise. The algorithm has one weight vector per class and the surrogate loss is a function of the linear activations (one per class). The surrogate loss of an example with linear activation vector a and class c has t…
New algorithm tackles risk-aware learning problems efficiently.
problem Risk-aware learning with mean-semideviation objective.
method Zeroth-order compositional stochastic optimization algorithm.
result Algorithm converges to optimal solutions with explicit rates.
This research improves PAC-Bayesian bounds for classification tasks using convexified loss.
problem Deriving generalization bounds for classification tasks with non-convex loss functions.
method Shift focus to misclassification excess risk bounds for PAC-Bayesian classification using convex surrogate loss and leveraging PAC-Bayesian relative bounds in expectation.
result Improved PAC-Bayesian bounds for classification tasks with convex surrogate loss.
Proposes a classifier with bounded abstention rate for binary classification.
problem Binary classification with abstention rate constraints.
method Characterizes Bayes optimal classifier, proposes plug-in classifier with abstention region, and develops computationally efficient algorithm.
result Proposed classifier achieves high probability of satisfying abstention constraint and is minimax near-optimal.
Proposes a method for inference in high-dimensional classification with non-differentiable surrogate losses.
problem Lack of inference procedures for identifying driving factors in high-dimensional classification with non-differentiable surrogate losses.
method Kernel-smoothed decorrelated score and cross-fitted version for hypothesis tests and interval estimators.
result Valid and superior inference methods for high-dimensional classification with non-differentiable surrogate losses.
We propose a class of very simple modifications of gradient descent and stochastic gradient descent. We show that when applied to a large variety of machine learning problems, ranging from logistic regression to deep neural nets, the proposed surrogates can dramatically reduce the variance, allow to take a larger step …
The area under the ROC curve (AUC) is a widely used performance measure in machine learning. Increasingly, however, in several applications, ranging from ranking to biometric screening to medicine, performance is measured not in terms of the full area under the ROC curve, but in terms of the \emph{partial} area under t…
Solves learning halfspaces with Massart noise for log-concave distributions.
problem Learning halfspaces with Massart noise in distribution-specific PAC model.
method Identifies a smooth non-convex surrogate loss and uses SGD to solve the learning problem.
result First computationally efficient algorithm for learning halfspaces with Massart noise for a broad family of distributions.
BMM algorithm improves convergence for nonconvex optimization problems.
problem Constrained nonsmooth nonconvex optimization problems.
method Block majorization-minimization with diminishing radius.
result Improved convergence rate for nonconvex optimization problems.
The paper proposes a method to model non-smooth functions using clustering, classification, and Gaussian process modeling.
problem Modeling discontinuities and non-smoothness in expensive computational models.
method Three-stage approach combining clustering, classification, and Gaussian process modeling.
result The approach successfully models discontinuities and non-smoothness in various functions.
Extends private optimization to non-convex problems efficiently.
problem Private optimization of non-convex functions over discrete and continuous domains.
method Two algorithms: one for discrete domains and one for continuous domains, both requiring boundedness and Lipschitz continuity.
result Oracle-efficient optimization algorithms for non-convex problems, outperforming standard approaches in some cases.
Consider a classification problem where we have both labeled and unlabeled data available. We show that for linear classifiers defined by convex margin-based surrogate losses that are decreasing, it is impossible to construct any semi-supervised approach that is able to guarantee an improvement over the supervised clas…
The Schatten-p norm (0<p<1) has been widely used to replace the nuclear norm for better approximating the rank function. However, existing methods are either 1) not scalable for large scale problems due to relying on singular value decomposition (SVD) in every iteration, or 2) specific to some p values, e.g., $1/…
Study on calibration and consistency of adversarial surrogate losses.
problem Designing robust classifiers with theoretical guarantees.
method Extensive analysis of H-calibration and H-consistency of adversarial surrogate losses.
result Some convex loss functions and supremum-based convex losses are not H-calibrated for important hypothesis sets.