On 27 November, Savantic hosted a breakfast seminar at their office. The theme of the talk was how Deep Learning technology has fundamentally changed image analysis. Claes Orsholm presented news both from the field of hardware and from the algorithm and model world – it will soon be possible to create advanced built-in Deep Learning models in small devices. New compression techniques make it possible to move the analysis processes, which, until now, have had to be carried out in large GPU based systems, to the hardware devices themselves. This means that neural networks can soon be an integrated part of smaller units like cars.
The biggest challenge is no longer to build the system itself, but to be able to create a working strategy for how to manage and analyse data, as well as knowing how to design the neural network. Claes Orsholm also talked about Deep Convolutional Neural Networks, which are more accurate than human experts, when it comes to image analysis. This fact, of course, will present a host of new possibilities for how such systems can be used.
Among the guests at this morning’s seminar were some 15 people representing the business sector. They were particularly interested in knowing more about the possibilities of training a neural network that is built into a small device and how to know when the device is ready to have its model validated, for instance in a car, with the rigid demands on safety it presents.
One conclusion from the presentation was that the tools needed for building Deep Learning features into individual devices are already a reality, but as regards data, a lot of work related to managing, structuring and validation remains to be done.