While there has been rapid progress in the adoption of neural networks and the evolution of neural network structures, the problem of training data remains. Even companies with access to the largest data sets still need additional data, and in particular corner case data. At the same time, many companies simply don’t have access to “big data” and need an alternative solution. The advent of powerful GPUs, able to give near photo realistic results in simulations, has led to the evolution of simulators for the creation of training and test data for AI systems. This talk will discuss the benefits of such an approach, as well as looking at the issues and limitations introduced by using artificial data. While this methodology has been driven by the automotive industry, its relevance to other industries such as surveillance and retail will be discussed.
Can Simulation Solve the Training Data Problem?
Session Speakers
-
Peter McGuinness
VP AI and Services, Mindtech Global
Peter has held various roles in the electronics industry, being closely aligned to graphics, AR/VR and AI throughout. Previously with Highwai as CEO, ST Microelectronics as Research Director and Imagination Technologies as the Director of Technology Marketing, Peter has recently joined Mindtech Global as VP AI and Services. Peter has many years’ experience in the field of AI, both from a technical and marketing point of view. He is currently Chairman of the Neural Network Exchange Format Working Group within Khronos Group. Peter has been working for multiple years on the creation of synthetic data for the training of neural networks and has personally dealt with many of the issues surrounding its generation and use. Peter is the author of 5 US patents.
