The growth of the performance of artificial intelligence (AI) technologies depends significantly on the availability of specialized chips. For example, many of the current AI systems use graphical processing units (GPUs) initially developed for computer games, especially by Nvidia. Intel has also developed recently new processors for AI, with radically improved technologies as compared to GPUs, inspired by the architecture of the brain. These neuromorphic chips represent information through spikes (pulses), in the same way as neurons in the brain.
A recent such chip from Intel, called Loihi, implements learning rules for spiking neural networks developed at the Romanian Institute of Science and Technology (RIST). These reinforcement learning rules allow the chip to learn from rewards and punishments, similarly to how animals are trained. More precisely, Loihi implements in hardware “synaptic eligibility traces”, components of a learning rule developed by RIST’s Răzvan Florian, which allow the implementation of this learning rule on the Loihi chip. The article describing this learning rule (R. V. Florian, Reinforcement learning through modulation of spike-timing-dependent synaptic plasticity, Neural Computation, 19/6, 2007) is among the 14 scientific publications cited by the scientific article describing Loihi. Intel’s scientists who created the processor have run on Loihi the reinforcement learning rule developed at RIST for solving a sequential decision-making problem.
The advantage of this type of neuromorphic chips is that they use up to 1000 times less energy than classical processors, which allows their efficient use for implementing AI capabilities in mobile devices such as phones or tablets.
Mordor Intelligence estimates that the global market size of neuromorphic chips is currently 1.4 billion dollars and that it will grow to 6 billion dollars in 2023, due to increase demand from recent AI developments.