To edge devices
High model optimization cost
Changes in device requirements will
AI services developed
High cloud operation costs
SubnetX is an AI model development support package that builds a vast number (tens of thousands or more) of trained deep learning models (subnets) with different computational complexity and accuracy by simply performing a predefined series of training on a certain deep learning model (Supernet), and achieves optimization by simply selecting the deep learning model that matches the target device. The package is an AI model development support package that can achieve optimization by simply selecting the appropriate deep learning model for the target device.
In our case study, we successfully reduced development time by 38% when optimizing the same model for three devices by leveraging a pre-trained SubnetX. We also confirmed that adapting a pre-trained SubnetX to a custom dataset speeds up adaptation to new datasets and has the potential to reduce EdgeAI Ops man-hours.
Normally, only one model can be developed in one cycle, resulting in repeated trial and error and rework whenever performance targets change. With scalable model development, tens of thousands or more diverse models can be acquired in a single cycle, minimizing rework and reducing development costs even if requirements change after the start of operation, during development, or after operation.
An application with AI model developed for iPhone13. However, the memory usage was too high and the UX was affected. As a result, the AI model had to be changed, and the development and operation costs continued to increase as the number of devices was increased or changed.
Industry: Automobile manufacturer
CHALLENGE: AI MODELS WERE DEVELOPED AHEAD OF TIME AND NEEDED TO BE OPTIMIZED WHEN DEPLOYED ACCORDING TO THE DEVICE SPECIFICATIONS OF THE OPERATIONAL SYSTEM.