Research Projects

A lot of computations are needed for a deep learning based generative model to learn from large datasets. Each computation cycle costs energy. As companies like Google transition their data centers to nuclear power to enhance sustainability and reduce reliance on fossil fuels, it becomes clear that the computational costs of AI are substantial. The energy-intensive nature of training and deploying large-scale models underscores the need for more energy-efficient AI solutions. This context reinforces the importance of research aimed at making AI models more sustainable. Sustainable AI aims to reduce energy consumption in two main areas: i) During Training, ii) Deployment. In this proposal, we focus on improving the energy efficiency during the deployment. What is Pruning: Pruning in machine learning involves removing unnecessary or less important parameters from a trained model to reduce its size and computational cost. Traditionally, this is done by evaluating the importance of each parameter based on criteria like weight magnitude or contribution to performance, then removing the unimportant ones. • To date, the pruning of deep generative models or Vision-Language Models (VLMs) using perceptual models has not been extensively addressed. The human visual system is highly efficient at processing visual information, and computational models such as the LGN (Lateral Geniculate Nucleus) and V1 (primary visual cortex) have shown notable parameter efficiency in visual data processing. • Building on these insights from the human visual system, this project aims to develop efficient Perceptual Quality preserving Pruning Frameworks for generative models, such as deep generative models and VLMs, without compromising perceptual quality. Vision-language modelling involves generating visual content, such as images, videos, and 3D models, from textual descriptions. It includes techniques like text-to-image generation (e.g., DALL-E3, Stable Diffusion) • Unlike traditional knowledge distillation methods, which often require costly and energy- intensive retraining, our approach focuses on pruning strategies that preserve key perceptual information without the need for retraining the models. • Our quality-preserving framework provides a hardware-independent solution for energy reduction, adaptable to any use case. It efficiently supports pre-trained models while ensuring high performance and broad compatibility across different use cases. In this proposal, we plan to make models more efficient and aim to minimize their carbon footprint, ensuring that AI technologies are more sustainable for real-world application