Review of DNN optimization and compression methods for Edge AI systems

During the last years, Deep Neural Networks (DNN) have gained a great importance in several research domains and applications that can learn from data. Thus, one can use DNN for image processing, actions recognition, text classification, language processing, speech recognition, etc. However, this high success of DNN models is accompanied by a high increase of computation time, architecture complexity, and memory storage. Moreover, the deployment and exploitation of DNN models in real environments is generally related to connected sensors (cameras, microphones, temperature sensors, etc.) where the inference process needs to be executed in real time on embedded or edge resources. In this work, we present an overview of the three main compression methods (pruning, quantization, knowledge distillation) that allow to reduce the computation time and memory size of DNN models while keeping a high accuracy. Experimental results are conducted within image classification, object detection and actions recognition with DNN models that were optimized and deployed on Edge AI resources such as Jetson Nano, Jetson Xavier and Jetson Orin. The obtained results showed that we could reach a compression factor of around 6x while keeping a good accuracy, and we also observe a speedup of 3x compared to non-compressed models.

With also:

• Nicolas Amand - "Intellectual property rights and database protection"
• David Lhoir - "The data management plan: concept, goals and tools"

• 2:00 pm: Thierry Dutoit - Welcome words
• 2:05 pm: Sidi Mahmoudi - "Review of DNN optimization and compression methods for Edge AI systems"
• 3:10 pm: Use cases demos
• 3:45 pm: Nicolas Amand - "Intellectual property rights and database protection"
• 4:00 pm: David Lhoir - "The data management plan: concept, goals and tools"
• 4:15 pm: End