Research

Computer Sciences and Information Technology

Title :

Study of Latent Space for Deep Learning based Image Generation Models using Topology

Area of research :

Computer Sciences and Information Technology

Focus area :

Computational Topology

Principal Investigator :

Dr. Nibaran Das, Jadavpur University, West Bengal

Timeline Start Year :

2024

Timeline End Year :

2027

Contact info :

Details

Executive Summary :

This project aims to uncover the latent space of generative models through mathematical modeling, addressing the weakness of many models in the generative space due to a lack of topological understanding. The project will provide extensive theoretical and mathematical proofs for modeling the latent space of a generative model based on model architecture and parameters. The use of computational topology, particularly Persistent Homology, will be employed to understand the topology of the latent space and infer mathematically how it came to be. This project could potentially solve some of the biggest problems in AGI and deep generative modelling. References to previous research include Wu et al.'s work on learning a probabilistic latent space of object shapes via 3D generative-adversarial modeling and Goodfellow's work ongenerative adversarial networks.This project might yield solutions to some of the biggest problems in AGI and deep generative modelling. References: [1] J. Wu, C. Zhang, T. Xue, W. T. Freeman, and J. B. Tenenbaum, “Learning a probabilistic latent space of object shapes via 3d generative-adversarial modeling,” 2017. [2] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020.

Total Budget (INR):

6,60,000

Organizations involved