Imagination Technologies
DEIB - BIO1 Room (Building 21 – First Floor)
October 31st, 2017
4.00 pm - 6.00 pm
Contacts:
Giovanni Agosta
Research Line:
System architectures
The evolution of artificial vision systems pushes for an integration of solutions with a high computational load on mobile/embedded systems, as for instance convolutional neural networks (CNN) for the real time video analysis.
The seminar introduces the basic concepts of CNN, their structure and their computational requirements. The speaker will also discuss more in detail the architecture of the PowerVR Series 2NX neural networks hardware accelerators, comparing them to other possible implementations of CNNs on hardware already present in mobile/embedded systems: CPU, DSP, GPU.
Finally, the speaker will talk about the dedicated tuning systems that allow the maintenance of a good level of accuracy on the networks by lowering the precision of the arithmetic operations used.