Novel Pruning and Optimization Technique on AI models reducing Computation cost of Training and Inference of AI Models

This ongoing research project focuses on developing novel pruning and optimization techniques to reduce the computational costs associated with training and deploying large AI models. The work addresses the growing concern of AI's environmental impact and accessibility limitations due to resource requirements.

Research Approach

The current investigation includes:

• Development of adaptive pruning algorithms that preserve model accuracy
• Implementation of quantization techniques for efficient model deployment
• Exploration of knowledge distillation to create lightweight models
• Benchmarking against state-of-the-art optimization methods

Preliminary Results

Initial findings indicate that the proposed techniques can reduce computational requirements by up to 70% while maintaining over 95% of the original model performance across several benchmark tasks in natural language processing and computer vision.