SCIENCE@home

Computational Creativity by Generative Adversial Network with Leaked Information

Sayantani Ghosh,Amit Konar,A. Nagar

Published 2024 in IEEE International Joint Conference on Neural Network

ABSTRACT

Computational creativity is defined as the ability of artificial systems to generate artifacts with substantial novelty and originality, comparable to those crafted by human experts. This study introduces an innovative approach to implementing computational creativity in the scientific domain, exemplified through the automatic generation of trigonometric identities using a novel LeakGAN model, a Generative Adversarial Network with leaked information. The novelty of the proposed LeakGAN lies in its discriminator model, constructed on the foundation of a Convolutional Neural Network (CNN), aimed at capturing the most relevant features from input data. These features are subsequently leaked to the generator model, enabling the production of identities with substantial originality and quality. The introduced novelty in the discriminator’s architecture encompasses the use of a new activation function called Mish, strategically employed to enhance the network's convergence rate and address over-fitting issues during training. Additionally, an attention layer is introduced to highlight the most relevant information within the feature space, thereby improving the network's learning capacity. Furthermore, a unique mixed pooling layer is utilized, combining the advantageous effects of max-pooling and average pooling operations to enhance the network's adaptability to varying feature distributions. Performance analysis, incorporating BiLingual Evaluation Understudy (BLEU) metrics and various comparative studies, substantiates the efficacy of the proposed LeakGAN model in generating novel identities compared to its traditional counterparts. Moreover, human evaluation involving 20 mathematical experts confirms the significant novelty of the generated identities compared to existing standard textbook problems. Consequently, the proposed technique proves valuable for generating diverse trigonometric identities suitable for inclusion as chapter-end exercises in middle school mathematics textbooks.

PUBLICATION RECORD

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-42 of 42 references · Page 1 of 1

CITED BY