Artificial intelligence (AI) models are being produced and used to solve a variety of business and technical problems. AI model engineering processes, platforms, and products are acquiring special significance across industry verticals. Due to deeper automation, the number of features being used for model generation is numerous and hence the resulting AI models are bulky. Therefore, AI researchers have come out with a number of powerful optimization techniques and tools to compress AI models. This paper explores a suite of pioneering methods for achieving AI model compression. Pruning and quantization techniques are being used together to reduce the size of complex AI model architectures and make them optimized and performant so that they can be easily deployed in the IoT edge devices. We have elaborated on three different methods to compress an AI model and drawn inferences on the same. Further, proper comparisons were conducted between model metrics on the basis of minimal loss in accuracy and precision. To identify the most accurate and compressed model, we utilized pruning techniques and dived deeper by comparing performances using right metrics. We have implemented two types of pruning for the purposes of this paper: weight pruning and unit pruning. These pruning techniques were implemented with 0 to 100 percent sparsity to create a network model that is lightweight and does not significantly affect performance metrics. We have also reduced the size of the model by removing weights and neurons that were used more than once. We have compared pruning and quantization compression methods, which are implemented using the Tensor flow lite library. We have found that Tensor flow lite model is better for edge deployment. Additionally, a practical survey of network compression techniques is also included.
Pruning and Quantization for Deeper Artificial Intelligence (AI) Model Optimization
Lect. Notes Electrical Eng.
International Conference on Robotics, Control, Automation and Artificial Intelligence ; 2022 November 24, 2022 - November 26, 2022
2023-11-18
13 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
British Library Online Contents | 2005
British Library Conference Proceedings | 2002
|