Turning towards the natural world for inspiration, many of humanity’s inventions and solutions are copied from nature. From burr-bearing plants – whose microscopic hooks led to the invention of velcro – to butterfly wings inspiring more efficient solar cells, it becomes ever more clear that solutions to some (if not all) of our challenges can be found in nature. In an effort to revolutionise and advance computing, engineers have also turned towards nature, using its most advanced “computer” to serve as a blueprint for neuromorphic computing – the human brain.
This may be the first you hear about neuromorphic computing, but the concept has been around since the late 1980s. Carver Mead, a renowned American engineer and scientist, first coined the term, which encompasses the use of very-large-scale integration (VLSI) systems with electronic analog circuits, creating artificial neural systems that mimic the human nervous system. Neuromorphic computing is thus inspired by the human brain’s architecture and its ability to learn from unstructured stimuli.
For those up to date with AI trends and emerging tech, neuromorphic computing might sound a lot like neural networks, but it can more accurately be understood as the bridge between traditional computer processors and the varying strengths and weaknesses of the human brain. Although modern CPUs and GPUs are capable of outperforming the brain with regards to serial processing tasks, they also consume vast amounts of energy. Thus, they are not able to match the low baseline energy footprint of the human mind. The complex task of reducing energy expenditure while also decreasing latency proved to be a roadblock for neuromorphic computing and, in order for the field of neuromorphic engineering to evolve and become feasible, a new type of hardware was needed: memristors.
For many years, the main passive electrical components were the capacitor, resistor and inductor. In 2008, a research team at HP Labs created the first memristor, a nanoscale circuit element capable of recalling past state electrical resistance values, making it usable as both memory or a processing unit. What makes memristors profound is that they are theoretically similar to brain synapses – in that they exhibit similar switching characteristics, which can be used to fabricate electronic synapses for neuromorphic computing products.
Neuromorphic computing systems rely on a combination of neuromorphic chips, which are crammed with artificial neurons and synapses. These neurons and synapses are able to mimic activity spikes within the human brain and handle all processing on the chip itself, resulting in smarter, energy-efficient systems with low latency. Neuromorphic chips also rely on spiking neural network (SNN) models to arrange internal elements in a manner that resembles neural networks within human brains. Each artificial neuron within the SNN is capable of firing independently, sending signals to other artificial neurons within the network, directly affecting electrical states. And by encoding information within the signals, SNNs can replicate natural learning processes.
Intel Labs, a global technology research organisation, developed a system, dubbed Loihi, that researchers could use to implement SNNs. Introduced in 2017, the Loihi chip encompasses 14nm process technology and its entire 128-core design is based on SNN algorithm architecture. Intel’s 5th generation neuromorphic research chip serves as a support system for SNNs and includes over 130 000 neurons, each of which can communicate with one another. Each one of Loihi’s 128 cores also contains learning engines, which developers can access programmatically to manipulate on-chip resources. The SNN-optimised architecture of the Loihi chip supports expedited and continuous learning in unorganised environments through programmable synaptic learning rules, while sporting low power consumption.
One of the most powerful neuromorphic systems to date is Intel’s Pohoiki Springs, made up of 768 Loihi chips. Packing 100 million artificial neurons while operating at a power level of under 500 watts, Pahoiki Springs is a massive breakthrough in the field of neuromorphic computing. But the processor specs still pale in comparison to the human brain, which sports over 86 billion neurons while operating at a mere 20 watt power level. The end goal of neuromorphic computing would be to match or exceed the energy efficiency of the human brain while still being able to learn and improvise within unstructured environments, but with Pohoiki Springs already on par with the brain of a small mammal, this generation might just live to witness the actualisation of this goal.
The concept of artificial systems that are able to learn from observations rather than programming gives way to a myriad of potential applications and future advancements within the field of technology. From driverless cars to domestic robots, the prospective applications of energy-efficient cognitive functions are endless. Even though neuromorphic chips have not quite reached the commercial stage, we simply can’t help but buzz with anticipation for the next generation of AI, which might just come to be in the not so distant future.