Issue link: https://htpgraphics.uberflip.com/i/1385717
LEAD RESEARCHER: LUIS PLANA I am one of the architects of the brain-inspired, million- processor SpiNNaker system, the world's largest neuromorphic supercomputer. SpiNNaker can support sophisticated and biologically-realistic models of substantial brain subsystems in real time, which allows it to interact with robots. The RAIN Hub is an ideal platform for the exploration of SpiNNaker-Robot interactions, focusing on vision processing and the visual cortex but also involving planning and motion control. The research is driven by two questions: Can massively-parallel computing resources accelerate our understanding of brain function? Can our growing understanding of brain function point the way to more efficient, parallel, fault- tolerant computing platforms? BIO-INSPIRED VISION FOR ENHANCED ROBOT SITUATIONAL AWARENESS SUMMARY // The human brain is extremely adept at object recognition, feature selection and adapting to changes in the environment. Recent significant advances in artificial neural nets for classification and natural language processing notwithstanding, no machine is close to matching the brain. These artificial models are costly to develop and train, both in monetary and environmental terms. Researchers have shown that training a neural net can be equivalent to five times the lifetime emissions of a car. Silicon retinas operate in a similar fashion to human eyes. They respond to light changes and send messages to the brain to highlight points of interest without any intervening computation. We are using these biologically-inspired sensors and spiking neural networks to create a vision system with potential advantages: reduced response time due to dynamic vision sensor operation, increased energy efficiency due to reduced computation and communication, and increased fault tolerance due to their resilience and redundancy. UNIQUENESS // Traditional computing platforms are designed around a somewhat small number of fast, very powerful processors. The human brain, on the other hand, consists of hundreds of billions of very simple, slow processors called neurons. Additionally, computers are optimised for the communication of large chunks of data while neurons only send very small messages. This surprising strategy lies at the heart of the energy efficiency of the brain. It also helps to explain how it can reach its immense scale. Our project uses a brain-inspired approach: SpiNNaker, the project "brain", incorporates over a million energy-efficient mobile-phone-class processors interconnected by a network that communicates with short messages, similar to those used in the brain. This architecture can simulate brain-like neural nets in real time, a key characteristic for interacting with robotic platforms. It can also scale-up to simulate the regions of the brain required for visual processing, planning and motor control. 48