HTP Graphics

RAIN PROGRESS // RAIN progress: The nature of the project requires effort along two different avenues: 1. A hardware interface between the SpiNNaker system and the sensors and actuators that will allow it to interact with the robotic platform, and 2. A Spiking Neural Network that can mimic the human visual cortex. Hardware development has progressed steadily and the current version of the FPGA-based interface can transport close to 50 Million visual events per second, enough to support more than one retina. This opens up the possibility of supporting binocular vision, with increased reliability and the capability for depth estimation. The development of the SNN, as expected, has proved to be the most challenging aspect of the project. Rate- based SNNs have not shown significant advantages over non-spiking ones, leading us to consider SNNs based on spatio-temporal encodings, currently at a much earlier stage of development. The software platform that allows the neural net to be mapped onto SpiNNaker and simulated in real time is ready and has been extensively tested. The technology will be demonstrated in collaboration with the Neurorobotics Laboratory at KTH Stockholm, using silicon retinas, a SpiNNaker machine and a robotic arm in an environment equipped with tracking cameras that can be used for verification. FUTURE ASPIRATIONS // Our longer-term goal is to support autonomous decision making in robotic platforms. Two lines of inquiry are of future interest: Extending the system capabilities to complete the biologically-inspired robotic control loop by incorporating SNNs that can mimic human planning and motor control, tasks usually associated with the Frontal Lobe and the Cerebellum. Our group is starting a collaboration, within the context of the Human Brain Project, for the development of a Spiking Neural Network model of the Cerebellum that can be run on SpiNNaker in real time. Similar collaborations will be sought for the development of motor control. Adding on-line learning capabilities to the system. SNNs open up the possibility of new learning algorithms. As opposed to artificial neural networks, SNNs can support unsupervised learning, a process in which the network learns in real time and by itself. Unsupervised learning has shown promise in optical character and pattern recognition. Incorporating unsupervised learning into our bio-inspired platform would increase its potential to support decision- making. Our extensive experience developing computing platforms indicates that processors are free, the real cost of computation is energy. The overarching goal, perhaps unattainable, is to match the energy efficiency and fault tolerance of the human brain. REMOTE INSPECTION 49

• Cover