[Paper Review] Reflection methods for user-friendly submodular optimization

Stefanie Jegelka, Francis Bach+1 more|arXiv (Cornell University)|Nov 18, 2013

Complexity and Algorithms in Graphs44 references66 citations

TL;DR

This paper introduces a novel reflection-based method for exact, user-friendly submodular function minimization by reformulating the problem as a continuous best approximation task. By leveraging the proximal formulation and solving it via reflection methods (e.g., Douglas-Rachford), the approach achieves fast convergence without hyperparameters, enables efficient parallelization, and outperforms existing methods in both speed and robustness on image segmentation tasks.

ABSTRACT

Recently, it has become evident that submodularity naturally captures widely occurring concepts in machine learning, signal processing and computer vision. Consequently, there is need for efficient optimization procedures for submodular functions, especially for minimization problems. While general submodular minimization is challenging, we propose a new method that exploits existing decomposability of submodular functions. In contrast to previous approaches, our method is neither approximate, nor impractical, nor does it need any cumbersome parameter tuning. Moreover, it is easy to implement and parallelize. A key component of our method is a formulation of the discrete submodular minimization problem as a continuous best approximation problem that is solved through a sequence of reflections, and its solution can be easily thresholded to obtain an optimal discrete solution. This method solves both the continuous and discrete formulations of the problem, and therefore has applications in learning, inference, and reconstruction. In our experiments, we illustrate the benefits of our method on two image segmentation tasks.

Motivation & Objective

• Address the practical inefficiency of existing submodular minimization algorithms, which are often slow or require complex parameter tuning despite being polynomial-time.
• Overcome limitations of subgradient and smoothing-based methods, which suffer from slow convergence and sensitivity to step-size choices.
• Develop a method that is both theoretically sound and practically efficient, enabling exact discrete solutions without relying on black-box or approximate solvers.
• Enable parallelization and easy implementation by exploiting the structure of decomposable submodular functions.
• Provide a unified framework that solves both continuous and discrete formulations of submodular minimization via a proximal problem with strong convexity.

Proposed method

• Reformulate discrete submodular minimization as a continuous best approximation problem using the proximal formulation: minimize $ f(x) + \frac{1}{2}\|x\|^2 $, where $ f $ is the Lovász extension of the submodular function.
• Apply reflection methods (e.g., Douglas-Rachford splitting) to solve the proximal problem, which avoids subgradient steps and hyperparameter tuning.
• Use the solution of the proximal problem to recover the optimal discrete solution via thresholding: $ S^* = \{ k \mid x^*_k \geq 0 \} $.
• Leverage dual decomposition and orthogonal projections to enable parallel computation across decomposable subfunctions.
• Work with the dual of the proximal problem to exploit smooth optimization techniques and ensure fast convergence.
• Use the duality gap (discrete and smooth) as a convergence metric, enabling early stopping without sacrificing optimality.

Experimental results

Research questions

• RQ1 Can reflection methods be effectively applied to submodular minimization to avoid hyperparameter tuning and improve convergence?
• RQ2 Does reformulating submodular minimization as a best approximation problem via the proximal formulation lead to faster and more robust optimization?
• RQ3 Can the proposed method achieve competitive performance compared to specialized solvers like Maxflow while remaining generic and parallelizable?
• RQ4 How does the convergence of the discrete duality gap compare to the smooth duality gap in practice?
• RQ5 To what extent can the method be scaled using parallelization without sacrificing solution quality?

Key findings

• The Douglas-Rachford (DR) reflection method converges significantly faster than both BCD and accelerated gradient methods, even outperforming state-of-the-art BCD on graph cut problems.
• The discrete duality gap shrinks faster than the smooth duality gap, indicating that high-accuracy solution of the smooth problem is not required to obtain the optimal discrete solution.
• For graph cut problems, the DR method achieves a 5-fold speedup using 8 cores, demonstrating strong parallelization efficiency.
• On average, the DR method is 2.55 seconds per image, which is 2–9 times slower than Maxflow, but this is remarkable given that DR solves the full regularization path and is generic and parallelizable.
• The method achieves exact discrete solutions without any hyperparameter tuning, unlike subgradient or smoothing-based approaches.
• Figure 1 shows that background noise in image segmentation disappears only when the duality gap is small, confirming that convergence to low duality gap ensures high-quality discrete solutions.