Intel introduces ‘self-learning’ neuromorphic test chip

Intel has announced a neuromorphic chip called Loihi that can “learn” based on feedback from its environment. The design is inspired by the workings of the human brain and Loihi is supposed to help speed up artificial intelligence calculations.

Intel built the chip around a mesh network of 130,000 cores that act as virtual neurons, each capable of communicating with thousands of other “neurons.” In this way, it can simulate networks of 130 million synapses. Each core can be programmed to adapt its ‘behaviour’ during the processing of the signals, for example to strengthen connections in the network. This is done based on the strength of incoming peaks in the signal and based on the frequency of those peaks.

This means that the network on the chip does not need to be trained, but becomes increasingly ‘smarter’ over time for tasks such as recognizing the movements of cars or bicycles. Loihi could do this significantly more efficiently than traditional systems used for neural networks. Intel produces the chip at 14nm and will make the test chip available to universities and research centers in the first half of 2018.

Intel does not disclose technical details about the chip’s components. In 2012, the company announced that it was experimenting with the use of lateral spin valves and memristors to mimic the functioning of the brain. If the somewhat obscure image is a representation of the chip design, the one with the neuromorphic network on the chip package would be combined with a traditional processor.

The computer industry has been working on new architectures for certain types of computation for some time. The traditional Von Neumann architecture, where data processing is based on central processors and memory, is not optimal for the algorithms used for artificial intelligence, such as those for recognizing objects based on large data sets.

One of the new architectures concerns neuromorphic computing, where chipmakers try to imitate the workings of the brain at the chip level. Networks on chips mimic the electrical system of the brain, in which signals are transmitted in the form of pulses and a higher frequency of electrical stimuli creates stronger connections. In addition to Intel, IBM and Qualcomm, among others, are working on chips based on this principle.