Vision AI when Data is Expensive and Constantly Changing
, Deep Learning Engineer, NVIDIA
, Product Manager, NVIDIA
Building vision AI applications to recognize huge numbers of classes or to work in constantly changing environments is a significant challenge for any developer. The need for extensive training datasets and frequent model refinement can derail even the best development teams. We’ll showcase how to quickly build and deploy applications for such scenarios, leveraging Metropolis Microservices & Reference Apps. We’ll focus on a retail self-checkout use case with typical challenges, like identifying many products with limited training data and refining the application as new products and packaging get added. We’ll explore how the system can adapt continuously with limited new data, often without retraining, by leveraging pretrained models, a self-labeled data pipeline, and a few-shot learning architecture.