Sony's MCT: Bridging AI research to real-time edge

The YOLO Vision 2023 (YV23) event, hosted at the Google for Startups Campus in Madrid, featured a lineup of speakers from across the AI community. Among them was Amir Servi, Sony’s Edge Deep Learning Product Manager, who gave a presentation on bridging the gap between AI research and the real-time edge, where he introduced Sony's Model Compression Toolkit (MCT).

Meet Amir Servi: Bridging research and real-time AI#

Amir's talk focused on model compression and quantization techniques tailored for efficient edge AI deployment.

Navigating the Challenges of Edge AI with MCT#

Amir discussed the challenges of deploying AI models on edge devices, emphasizing the hurdles posed by limited resources and hardware restrictions. During his talk, he introduced Sony's Model Compression Toolkit (MCT), an open-source tool that works with PyTorch and TensorFlow models. MCT also powers the Ultralytics Sony IMX500 export.

Unlocking MCT's Potential#

Amir walked through MCT's key features, from hardware-aware quantization to state-of-the-art algorithms and automated parameter search, which make it a versatile toolkit for real-world AI deployment.

Fig 1. Amir Servi presenting at YOLO Vision 2023 at the Google for Startups Campus in Madrid.

Quantization techniques demystified: Results speak louder#

Amir explained quantization techniques such as post-training quantization (PTQ) and GPTQ and shared their results, including strong accuracy from PTQ with mixed precision and high compression rates for the Ultralytics YOLOv8 model.

Wrapping up#

Amir’s talk showed how tools like MCT help bridge AI research and real-time implementation, especially for YOLO models running on edge devices.

Curious to learn more? Watch the full Sony MCT talk.