Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Generate more than one child in your co-evolutionary semi-supervised learning GAN

Francisco Sedeño, Jamal Toutouh, Francisco Chicano

TL;DR

The paper addresses SSL-GAN training with limited labeled data by proposing a co-evolutionary approach, CE-SSLGAN, featuring panmictic populations, elitist replacement, and multiple offspring per generation ($\\lambda>1$). It couples two neural populations (generators $G$ and discriminators $D$) and evolves them under a $(\\mu+\\lambda)$ scheme while training offspring for $n_t$ epochs within budget $T_B$. Empirical results on RING, BLOB, and MNIST demonstrate that CE-SSLGAN generally outperforms standard SSL-GAN, particularly when using multiple offspring and extended training per generation, with MNIST showing improved generation and discrimination metrics after longer training. The work provides practical guidelines for hyper-parameter tuning and highlights potential for parallelization and multi-objective extensions in SSL-GAN frameworks.

Abstract

Generative Adversarial Networks (GANs) are very useful methods to address semi-supervised learning (SSL) datasets, thanks to their ability to generate samples similar to real data. This approach, called SSL-GAN has attracted many researchers in the last decade. Evolutionary algorithms have been used to guide the evolution and training of SSL-GANs with great success. In particular, several co-evolutionary approaches have been applied where the two networks of a GAN (the generator and the discriminator) are evolved in separate populations. The co-evolutionary approaches published to date assume some spatial structure of the populations, based on the ideas of cellular evolutionary algorithms. They also create one single individual per generation and follow a generational replacement strategy in the evolution. In this paper, we re-consider those algorithmic design decisions and propose a new co-evolutionary approach, called Co-evolutionary Elitist SSL-GAN (CE-SSLGAN), with panmictic population, elitist replacement, and more than one individual in the offspring. We evaluate the performance of our proposed method using three standard benchmark datasets. The results show that creating more than one offspring per population and using elitism improves the results in comparison with a classical SSL-GAN.

Generate more than one child in your co-evolutionary semi-supervised learning GAN

TL;DR

The paper addresses SSL-GAN training with limited labeled data by proposing a co-evolutionary approach, CE-SSLGAN, featuring panmictic populations, elitist replacement, and multiple offspring per generation (). It couples two neural populations (generators and discriminators ) and evolves them under a scheme while training offspring for epochs within budget . Empirical results on RING, BLOB, and MNIST demonstrate that CE-SSLGAN generally outperforms standard SSL-GAN, particularly when using multiple offspring and extended training per generation, with MNIST showing improved generation and discrimination metrics after longer training. The work provides practical guidelines for hyper-parameter tuning and highlights potential for parallelization and multi-objective extensions in SSL-GAN frameworks.

Abstract

Generative Adversarial Networks (GANs) are very useful methods to address semi-supervised learning (SSL) datasets, thanks to their ability to generate samples similar to real data. This approach, called SSL-GAN has attracted many researchers in the last decade. Evolutionary algorithms have been used to guide the evolution and training of SSL-GANs with great success. In particular, several co-evolutionary approaches have been applied where the two networks of a GAN (the generator and the discriminator) are evolved in separate populations. The co-evolutionary approaches published to date assume some spatial structure of the populations, based on the ideas of cellular evolutionary algorithms. They also create one single individual per generation and follow a generational replacement strategy in the evolution. In this paper, we re-consider those algorithmic design decisions and propose a new co-evolutionary approach, called Co-evolutionary Elitist SSL-GAN (CE-SSLGAN), with panmictic population, elitist replacement, and more than one individual in the offspring. We evaluate the performance of our proposed method using three standard benchmark datasets. The results show that creating more than one offspring per population and using elitism improves the results in comparison with a classical SSL-GAN.
Paper Structure (10 sections, 6 equations, 9 figures, 4 tables, 2 algorithms)

This paper contains 10 sections, 6 equations, 9 figures, 4 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. SSL-GAN training
  4. Co-evolutionary Elitist SSL-GAN
  5. Experimental setup
  6. Experimental results
  7. RING experimental results
  8. BLOB experimental results
  9. MNIST experimental results
  10. Conclusions and future work

Figures (9)

  • Figure 1: Example of SSL-GAN for MNIST digits generation
  • Figure 2: Representation of the datasets applied in our empirical analysis
  • Figure 3: Influence of $n_t$ on RING dataset
  • Figure 4: Influence of population and offspring size in the performance for RING
  • Figure 5: Evolution of the classification accuracy and $W_D$ through each training epoch performed for the selected methods on RING (first 300 training epochs)
  • ...and 4 more figures