Goal and Objectives or brief
Our research focuses is to optimize AI models for resource-constrained edge
platforms by addressing redundancies and applying probabilistic and statistical methods to model
parameters. This approach aims to design computationally optimized and memory-efficient models
for vision-based tasks. Our work involves model optimization methods such as pruning,
quantization, hardware-specific optimization, and operator and model fusion to streamline models
while preserving accuracy. Currently, we are also exploring redundancies in LLM encoders and
decoders for image denoising and text-to-image applications, with a focus on embedding and
attention matrix optimization. We investigate multi-objective Neural Architecture Search (NAS),
enhanced by genetic algorithms to ensure optimal model performance for edge devices.