Microsoft has unveiled that HoloLens 2 will feature an AI coprocessor, at the Computer Vision and Pattern Recognition (CVPR) 2017 event. Harry Shum, the executive VP of Microsoft's AI and Research Group, explained how this feature of HoloLens HPU 2.0 will help with the implementation of DNNs.
The HoloLens HPU (Holographic Processing Unit) is a custom-made coprocessor, designed for Microsoft HoloLens. Microsoft has been working on the second version of HPU for some time, and Shum unveiled the first secret of HoloLens 2.
Thanks to a new AI coprocessor, HoloLens will be able to natively and flexibly implement Deep Neural Networks. The chip will support a variety of layer types, fully programmable by Microsoft.
The AI coprocessor can run continuously, off the HoloLens battery, allowing features such as sophisticated hand tracking with hand-part segmentation. In addition, it will offer on-device voice recognition without sending samples to the cloud.
HoloLens 2 and DNNs
It is safe to say that deep learning has taken the world of recognition tasks by storm. Thanks to Deep Neural Networks we have seen large improvements in the accuracy of recognition.
However, deep learning approaches are challenging because they require large amounts of data for training. In addition, they require a type of compute that is not possible in current -general purpose- processor/memory architectures.
With that in mind, Microsoft has created an AI processor developed to face up to these challenges. This means that HoloLens 2 will analyze what the user sees and hears on the device, instead of sending data to the cloud for processing.