[Paper Review] Interaction Matters: A Note on Non-asymptotic Local Convergence of Generative Adversarial Networks

Tengyuan Liang, James Stokes|arXiv (Cornell University)|Feb 16, 2018

Mathematical Biology Tumor GrowthMathematics95 citations

TL;DR

This paper presents a non-asymptotic local convergence theory for smooth two-player games, showing how off-diagonal interaction affects SGA and how four stabilized dynamics (OMD, CO, IU, PM) achieve exponential convergence in unstable cases, with connections and learning-rate guidance.

ABSTRACT

Motivated by the pursuit of a systematic computational and algorithmic understanding of Generative Adversarial Networks (GANs), we present a simple yet unified non-asymptotic local convergence theory for smooth two-player games, which subsumes several discrete-time gradient-based saddle point dynamics. The analysis reveals the surprising nature of the off-diagonal interaction term as both a blessing and a curse. On the one hand, this interaction term explains the origin of the slow-down effect in the convergence of Simultaneous Gradient Ascent (SGA) to stable Nash equilibria. On the other hand, for the unstable equilibria, exponential convergence can be proved thanks to the interaction term, for four modified dynamics proposed to stabilize GAN training: Optimistic Mirror Descent (OMD), Consensus Optimization (CO), Implicit Updates (IU) and Predictive Method (PM). The analysis uncovers the intimate connections among these stabilizing techniques, and provides detailed characterization on the choice of learning rate. As a by-product, we present a new analysis for OMD proposed in Daskalakis, Ilyas, Syrgkanis, and Zeng [2017] with improved rates.

Motivation & Objective

• Motivate a non-asymptotic, local convergence understanding for GANs as two-player zero-sum games.
• Characterize how the off-diagonal interaction term influences convergence speed under SGA.
• Develop and unify analysis for stabilized saddle-point dynamics (OMD, CO, IU, PM) and provide learning-rate guidance.
• Connect stabilized methods to the curvature induced by interaction terms in bilinear/unstable local settings.

Proposed method

• Formulate GAN training as a smooth two-player zero-sum game with a local Nash equilibrium.
• Derive non-asymptotic exponential convergence results for SGA under local strong convexity-concavity (Theorem 1).
• Isolate the slow-down caused by the interaction term C Introduce matrix F and parameters alpha, beta for rate characterization.
• Analyze unstable (bi-linear) local games and show exponential convergence of four stabilized dynamics (Theorems 3–6).
• Provide explicit learning-rate choices for each method in the bi-linear setting and relate methods to interaction curvature.
• Offer improved analysis for Optimistic Mirror Descent (OMD) compared to prior results.

Experimental results

Research questions

• RQ1 How does the off-diagonal interaction term in two-player games affect non-asymptotic convergence rates?
• RQ2 Can standard gradient dynamics (SGA) be shown to converge exponentially to stable local Nash equilibria with an appropriate learning rate?
• RQ3 Do stabilized saddle-point dynamics (OMD, CO, IU, PM) yield exponential convergence for unstable/local bilinear games, and how are learning rates chosen?
• RQ4 What are the connections among these stabilized methods and how do they leverage interaction curvature to stabilize GAN training?
• RQ5 How do these local results relate to practical GAN objective functions and what guidance do they provide for learning-rate selection?

Key findings

• SGA converges exponentially to a stable local Nash equilibrium with a carefully chosen fixed learning rate, but slow-down is caused by the off-diagonal interaction term.
• In unstable or near-bi-linear local games, SGA diverges for any non-zero learning rate, while four stabilized dynamics achieve exponential convergence (OMD, CO, IU, PM).
• The interaction term curvatures are leveraged similarly across stabilized methods to stabilize dynamics and achieve faster convergence in the bilinear setting.
• A unified perspective links OMD, PM, CO, and implicit updates, showing they share curvature-utilization mechanisms for stabilization.
• The paper provides explicit learning-rate guidance for these methods in the bi-linear (unstable) setting and contrasts their performance with standard SGA.