Computers learn to learn

  

Intel and researchers from Heidelberg and Dresden will present hardware prototypes of the BrainScaleS, SpiNNaker and Loihi chips during the NICE Workshop, promising that they will have a major impact on the future of AI.

BrainScaleS is said to have a mixed analogue and digital design and works 1,000 to 10,000 times faster than real time.

The researchers have claimed that the second generation neuromorphic BrainScaleS chip has freely programmable on-chip learning functions, as well as an analogue hardware model of complex neurons with active dendritic trees. These, they explained, are especially valuable for reproducing the continual process of learning.

The second generation of SpiNNaker is apparently based on many-core architecture, which was developed by professor Dr Steve Furber at the University of Manchester.

The new SpiNNaker chip is based on so-called ARM architecture; a multitude of processor cores has been integrated on this chip. The team claimed that a single chip contains 144 ARM Cortex M4 cores, with innovative power management for efficient energy usage.

The chip is also designed to deliver computational power of 36 billion instructions per second, per watt. Its primary application with be for real-time simulation of multi-scale brain models.

Intel said that the Loihi research chip contains a highly developed command structure for neural networks from ‘firing’ neurons, as well as microcode-programmable learning rules. It is designed to support a range of on-chip learning models from supervised and unsupervised learning, to reinforcement-based learning processes.