Data is the core of any machine learning application. To make a machine learning model recognize patterns, objects, or behaviours it needs to be trained on large amounts of data that represent what the model is supposed to learn. For computer vision tasks this training data often consists of photos and videos gathered from the real world. However, using real-world data to train neural networks comes with several limitations and challenges with regards to quality, flexibility, and privacy. In this presentation Agnes Jernström of Neonode will explain what Synthetic Data is and how it can be used to overcome these challenges.

The presentation will cover:
The importance of high-quality training data for neural networks. Examples from the recent regulatory framework for AI presented by the European Commission [1], research on bias in neural networks by J. Buolamwini [2], and a study performed by MIT researchers of annotation errors in common computer vision datasets [3].
Introduction to synthetic training data with examples from how we have used synthetic data when developing our driver monitoring system at Neonode.
How synthetic data can be used to improve an AI solution by:
Making it possible for neural networks to learn things that are hard or impossible to annotate in real-world data.
Simplifying the creation of unbiased datasets
Avoiding the potential legal issues of using real human faces in the training data
Making it possible to tailor the training data to the task you are solving with regards to image size, camera position and lens parameters.