Source-linked AI summary

Accelerated Proximal Stochastic Dual Coordinate Ascent for Regularized Loss Minimization

Shai Shalev-Shwartz, Tong Zhang

arXiv:1309.2375v2stat.MLcs.LGmath.NAstat.CO

TL;DR

The paper addresses efficient regularized loss minimization for linear predictors, including settings with broad regularizer and loss-function structures. It introduces proximal SDCA and accelerates it through strongly regularized inner problems. The resulting runtime improves state-of-the-art results across several machine-learning optimization problems, while the acceleration procedure remains tied to Euclidean-strongly-convex regularizers.

  • Problem

    The paper studies how to efficiently solve regularized loss minimization for linear predictors across convex losses and regularizers.

  • Method

    It combines proximal stochastic dual coordinate ascent with an inner-outer acceleration procedure that solves a sequence of more strongly regularized problems.

  • Results

    The resulting method improves state-of-the-art runtime results for several machine-learning optimization problems, including SVM, logistic regression, ridge regression, Lasso, and multiclass SVM.

  • Takeaways & Limitations

    The framework supports general strongly convex regularizers and vector-valued losses, extending proximal SDCA to multiclass and nonsmooth-loss settings.

  • Takeaways & Limitations

    The acceleration procedure is designed for regularizers strongly convex with respect to the Euclidean norm, leaving extension to arbitrary norms open.

Abstract

from arXiv · show

We introduce a proximal version of the stochastic dual coordinate ascent method and show how to accelerate the method using an inner-outer iteration procedure. We analyze the runtime of the framework and obtain rates that improve state-of-the-art results for various key machine learning optimization problems including SVM, logistic regression, ridge regression, Lasso, and multiclass SVM. Experiments validate our theoretical findings.

1 Introduction

The paper develops a proximal stochastic dual coordinate ascent method and accelerates it through an inner-outer procedure, targeting regularized loss minimization across broad problem classes. Its runtime improves prior results, including for SVM, logistic regression, ridge regression, Lasso, and multiclass prediction.

  • The framework solves regularized loss minimization for linear predictors with convex losses, convex regularizers, and a regularization parameter.
  • When the condition number 1/(λγ) is O(n), the runtime is nearly linear in the data size, ˜O(dn).
  • The method improves state-of-the-art runtime results across key applications, including SVM, logistic regression, ridge regression, Lasso, and multiclass SVM.
  • The proximal SDCA method generalizes prior analysis to arbitrary strongly convex regularizers and vector-valued losses smooth or Lipschitz under general norms.
  • The acceleration procedure repeatedly approximately minimizes objectives with stronger regularization, centered around vectors from previous iterates.
  • The analysis adapts Nesterov’s estimation-sequence technique to approximate and stochastic proximal mappings, while allowing multiple passes over the data to avoid polynomial dependence on 1/ϵ.

2 Preliminaries

The preliminaries establish the convex-analytic setting, norm and smoothness conventions, and the primal–dual relationship used throughout the paper.

  • All functions are proper convex functions over a Euclidean space, with extended-real values allowed for function ranges.
  • The paper defines conjugate functions, dual norms, standard L2, L1, and L∞ norms, and norm-induced matrix operator norms.
  • L-Lipschitzness, (1/γ)-smoothness, and γ-strong convexity are defined relative to norms and their associated dual norms.
  • The primal objective has a dual formulation, and strong convexity of g implies differentiability of its conjugate g∗.
  • Primal–dual optimal solutions share the same objective value, while the duality gap upper-bounds both primal and dual sub-optimality.

3 Main Results

The paper develops Prox-SDCA for a broad regularized loss-minimization problem and accelerates it through an inner–outer procedure. The analysis covers smooth and Lipschitz losses, with runtime guarantees for different condition-number regimes.

  • Prox-SDCA: Prox-SDCA generalizes stochastic dual coordinate ascent to strongly convex regularizers and vector-valued losses smooth or Lipschitz under general norms.This includes non-smooth regularizers such as L1 regularization.
  • Prox-SDCA: At each iteration, the method selects one training example uniformly at random and updates only its associated dual vector.The remaining dual vectors stay unchanged.
  • Prox-SDCA: The proximal dual update replaces direct maximization with a lower-bounding proximal objective when the conjugate regularizer is difficult to optimize.Alternative gradient-based updates are also analyzed for complex loss conjugates.
  • Acceleration: The accelerated procedure improves dependence on the condition number when R2/(λγ) is much larger than n, under a Euclidean strong-convexity assumption.Generalization to strongly convex regularizers under general norms is left for future work.
  • Acceleration: Acceleration iteratively minimizes objectives with stronger regularization centered at a momentum point formed from previous iterates.Each outer iteration calls Prox-SDCA on a modified objective containing a quadratic proximal term.
  • Lipschitz losses: For non-smooth, non-differentiable losses that are L-Lipschitz, the paper applies smoothing so the resulting conjugate has bounded domain and the smoothed loss is smooth.The conjugate of an L-Lipschitz loss is infinite outside the dual-norm ball of radius L.

4 Proof of Theorem 3

The proof of acceleration combines approximate proximal minimization with lower-bounding quadratic functions and an estimation-sequence argument. It establishes both geometric error decay and bounded runtime for the inner Prox-SDCA calls.

  • Lower bounds: An approximate minimizer of the strongly regularized objective yields a quadratic lower bound on the original objective.This lower bound is the bridge from inner proximal solves to progress on P.
  • Estimation sequence: The algorithm maintains convex combinations of quadratic lower bounds, producing functions h_t that lower-bound P everywhere.The induction uses the recurrence defining h_t and the proximal lower-bound lemma.
  • Estimation sequence: The proof controls the gap between P(w(t)) and h_t(v(t)) by an error term ξ_t that contracts geometrically.The stated contraction is ξ_t ≤ ξ_1 e^{-(t−1)η/2}.
  • Convergence: Choosing η and the momentum centers y(t) appropriately makes the stopping condition imply P(w(t))−P(w*)≤ϵ after O(η^-1 log(ξ_1/ϵ)) iterations.The proof uses ξ_t ≤ e^{-η(t−1)/2}ξ_1 to derive the iteration requirement.
  • Runtime: Each outer iteration approximately minimizes a strongly regularized objective, and its averaged cost is bounded through the Prox-SDCA runtime result.The proof explicitly returns to the modified objective used by the accelerated procedure.

5 Applications

The paper instantiates its framework for common losses and regularizers, deriving Prox-SDCA and accelerated runtimes for ridge regression, logistic regression, Lasso, SVM, and related problems.

  • Application framework: The framework covers ridge regression, SVM, Lasso, logistic regression, and multiclass prediction through specific losses, conjugates, and regularizers.The application section describes these models and their algorithmic specializations.
  • Loss functions: Smooth hinge loss is (1/γ)-smooth, while the smoothed max-of-hinge loss supports multiclass prediction.The smoothed hinge construction preserves a controlled approximation to vanilla hinge loss.
  • Regularization: L1-L2 regularization adds slight L2 curvature to the L1 norm, enabling the framework to handle otherwise non-strongly-convex regularization.For sufficiently small λ, the resulting objective approximates σ∥w∥1.
  • Ridge regression: For ridge regression, Prox-SDCA has a runtime depending on R2/λ, while acceleration improves the runtime when R2/λ is much larger than n.Here R = max_i ∥x_i∥.
  • Lasso: For Lasso, the method provides runtime bounds parameterized by either σ and the optimal solution scale or B = ∥w̄∥2, and improves a prior bound by at least √n.The bounds use λ = ϵ(σ/ȳ)2 or λ = ϵ/B2 under the stated assumptions.
  • SVM: For SVM, choosing γ = ϵ yields an ϵ-accurate solution, with the accelerated rate reported as better than SGD in the stated regime.The comparison is expressed through the condition involving λϵ and n.

6 Experiments

Experiments compare Prox-SDCA, Accelerated-Prox-SDCA, and FISTA on three binary-classification datasets across several L1-L2 regularization settings. Accelerated-Prox-SDCA is consistently fastest and often significantly outperforms the alternatives.

  • Experimental setup: The experiments compare Prox-SDCA, Accelerated-Prox-SDCA, and FISTA on three large, diverse datasets.The datasets are astro-ph, CCAT, and cov1, with differing feature counts and sparsity.
  • Experimental setup: The binary-classification experiments use normalized features, smooth hinge loss with γ = 1, and σ = 10^-5.λ varies over {10^-6, 10^-7, 10^-8, 10^-9}.
  • Evaluation: The plots report primal objective versus passes over the training data, with methods stopped at guaranteed accuracy 10^-3.FISTA uses one pass per iteration, while n coordinate updates correspond to one pass for the SDCA methods.
  • Results: Accelerated-Prox-SDCA yields the best results and often significantly outperforms Prox-SDCA and FISTA.Prox-SDCA slows substantially when λ is small, while the relative performance of FISTA and Prox-SDCA depends on λ and n.

7 Discussion and Open Problems

The paper concludes that its proximal and accelerated SDCA framework improves state-of-the-art runtimes in many cases, while identifying two unresolved extensions.

  • Conclusion: The resulting method improves state-of-the-art runtime results in many machine-learning optimization cases.The conclusion follows the paper’s proximal SDCA method and its acceleration procedure.
  • Open problems: When 1/(λγ) is larger than n, the authors leave open whether the accelerated procedure can achieve a better runtime.This is identified as the first main open problem.
  • Open problems: The acceleration procedure is currently designed for Euclidean strongly convex regularizers, leaving extension to general norms open.The underlying Prox-SDCA procedure and analysis already support strongly convex regularizers with respect to arbitrary norms.

A Proofs of Iteration Bounds for Prox-SDCA

The proofs establish iteration and accuracy guarantees for randomized Prox-SDCA by combining coordinate-wise dual progress, norm-based smoothness, and high-probability conversion arguments.

  • Proof strategy: The proof generalizes prior SDCA analysis to strongly convex regularizers and loss smoothness or Lipschitzness under general norms.The argument follows Shalev-Shwartz and Zhang while extending the handled setting.
  • Prox-SDCA analysis: Each randomized iteration updates one dual vector selected uniformly at random, leaving the remaining dual vectors unchanged.The analysis conditions on the current dual iterate and bounds expected progress.
  • Prox-SDCA analysis: The key progress argument uses smoothness of the regularizer conjugate and strong convexity of the loss conjugate.These properties control the dual improvement under a coordinate update.
  • Accuracy guarantees: The theorems convert expected dual progress into primal accuracy and duality-gap guarantees through averaging, random selection, Markov’s inequality, and repetition.Repeating logarithmically many trials yields probability at least 1 − δ.
Loading 1309.2375v2…