Sharon is a Mechatronic Engineer passionate about applying machine learning, computer vision and robotics to complex world challenges. At CSIRO she tested and analysed sensor data for automated 3D imaging systems with space and manufacturing applications. At Nearmap she currently is developing a method to leverage domain knowledge to improve deep learning models for aerial imagery.
Talks and Events
2022 Talk: Leveraging Domain Knowledge for Deep Learning based Computer Vision
Deep learning models require massive amounts of data to be perform accurately. As the world is inherently interconnected, we can leverage relationships amongst identifiable objects to improve Deep Learning. For example, a shingle roof can not be tile roof, but both are roofs. So, we set ourselves a challenge: “How can we leverage the knowledge and the relationship amongst the things we see in our world to improve our data, software systems and the deep learning model?”. In this presentation, we share our experiences with knowledge graphs as a technique to model domain knowledge and reason about it to derive an embedding. We have leveraged these embeddings in numerous applications to build a more scalable, reliable and efficient AI System. The applications includes improving quality and richness of our datasets, identifying gaps in annotators’ knowledge, utilising existing data to synthesis new objects on the fly, and also to increase efficiencies of deep learning models.