- Currently thesis under progress which explores a novel form of Generative Model inspired by Variational Autoencoders (VAEs). The key contribution is a modified recognition model that operates on the principle of stochastic mixtures of Gaussians rather than producing a single Gaussian for each input. The study focuses on employing Stochastic Neural Networks as the recognition model within a VAE framework.
- While traditional VAEs are trained to maximize a lower bound on the log likelihood of the data (commonly referred to as the Evidence Lower Bound or ELBO), this research aims to achieve a tighter bound by incorporating stochasticity:
$$\text{Log Likelihood} \geq \text{ELBO}_{\text{stochastic}} \geq \text{ELBO}_{\text{VAE}}$$
- Explored techniques to enhance the performance of Variational Autoencoders (VAEs), focusing on improving generative modeling through advanced methods such as Importance Weighted Autoencoders (IWAE) and Beta-VAE. This work bridges the gap between theoretical foundations and practical applications of VAEs across diverse domains.
- Researched advanced Generative Adversarial Network (GAN) techniques, including Wasserstein GANs, Conditional GANs, and Laplacian Pyramid GANs, to generate high-resolution and high-fidelity images. Addressed key challenges such as mode collapse and artifacts in generative outputs.
- Conducted an in-depth study of Denoising Diffusion Probabilistic Models (DDPMs), delving into their mathematical foundations. Further explored state-of-the-art diffusion techniques, including the Latent Diffusion Model, to advance the field of generative modeling.
- Developed a strong foundation in probability theory, emphasizing its application in probabilistic models such as VAEs and DDPMs. This understanding has been instrumental in analyzing uncertainty, optimizing model performance, and interpreting complex generative processes.
- Advanced Computer Vision Research in Sports Vision domain to develop a real-time offside detection system for soccer matches using a single broadcast camera. The system leverages computer vision techniques to detect offside events with high accuracy, reducing the dependency on the 10-30 cameras used in existing systems.