Paper Analysis - Generative Minimization Networks: Training GANs Without Competition

[NicolaBernini]

Analysis of the paper

Generative Minimization Networks: Training GANs Without Competition

https://arxiv.org/abs/2103.12685v1

[NicolaBernini]

Problem Description

GANs require solving an optimization problem which is harder than the single objective optimization problem typically observed in current ML: it is a MinMax optimization problem, so it combines multiple objectives, the ones of the 2 adversarial players

Using the usual tools, i.e. gradient based methods, works in theory but practically there are problems in terms of convergence:

• it may never happens (cycle) or
• be suboptimal (so takes many iterations and training takes super long to converge)

Solution

• The paper uses the theory of game theory to move away from the problematic MinMax objective and use another one which gives better theoretical guarantees and it is easier to solve