Neuromorphic Computing Rethinks AI Chip Design
One of the classic movie scenes of all time is when Gene Wilder’s reluctant mad scientist, Dr. Frederick Frankenstein (Franken-STEEN!), realizes his incompetent henchman Igor got his hands on an abnormal brain (or, as the hunchback believed, it belonged to someone named Abby Normal). Hilarity ensued. However, that cinematic moment also reminds us that if you use subpar hardware, the results will be less than expected. The emergence of artificial technology as a disruptive technology in everything from healthcare to fintech has largely been a revolution in software and big data, such as developing algorithms that can recognize faces or analyze medical images. Hardware has largely been an afterthought. Even Nvidia’s (NVDA) powerful parallel Graphical Processing Units (GPUs) for deep learning and other AI applications were originally designed for other purposes such as gaming. While a number of companies are now developing hardware specifically for AI, some believe neuromorphic computing is the future.
What is Neuromorphic Computing?
To understand the basics of neuromorphic computing – not an easy task – it’s worthwhile to discuss a few essentials about computer hardware without getting too bogged down in terminology about transistors or Tensor Processing Units (TPUs).
Let’s start with Moore’s Law, a concept introduced in the 1960s by Intel (INTC) co-founder Gordon Moore that says computers will get faster and better about every two years, with a corresponding drop in price for the technology. The principle is based on the Central Processing Unit (CPU), which has served as the brain of most computers since Intel introduced the single-core CPU in the 1970s. Moore’s Law has mostly worked as advertised, but the end is nigh, as many experts predict that it won’t be possible for much longer to cram enough electrical components into a circuit and expect even better performance. That’s led some to bet on moonshot technologies like quantum computers as a way to radically break free of Moore’s Law, though whether the “quantum advantage” will ever arrive remains to be seen.
Meanwhile, many AI chip companies – including a cohort out of China, not to mention Google and its TPU – are doing their own riffs on the GPU in order to build faster and more energy-efficient hardware that improves on the design, which relies on multiple cores for parallel computations, as opposed to the handful of processing cores used by CPU systems. Part of the motivation is the trend toward mobile and edge computing, where the emphasis is on giving devices enough “intelligence” to work with minimal need to tap into computing power supplied remotely through the cloud (though superfast 5G could theoretically assuage concerns over latency).
One emerging technology that could tick many boxes for better AI hardware – low latency, low power – is neuromorphic computing. Caltech scientist Carver Mead coined the term back in the 1980s to describe a sort of hybrid analog and digital circuit. Neuromorphic computing or engineering is a type of biomimicry that attempts to build a computer system that mimics the biological structure of the human nervous system. More specifically, “neuromorphic chips attempt to model in silicon the massively parallel way the brain processes information as billions of neurons and trillions of synapses respond to sensory inputs such as visual and auditory stimuli.” That quote comes from an article in MIT Technology Review on Qualcomm’s (QCOM) efforts to develop the technology back in 2014.
Spiking Neural Network
Much of the work around neuromorphic computing involves what’s called a Spiking Neural Network (SNN), which is modeled on how neurons interact with synapses, which send signals between brain cells chemically or electrically. ZDNet did a great deep dive into neuromorphic computing and sums up the SNN pretty well:
In the biological brain, each neuron is connected to a variety of inputs. Some inputs produce excitation in the neuron, while others inhibit it – like the positive and negative weights in an artificial neural net. But with an SNN, upon reaching a certain threshold state described by a variable (or perhaps with a function), the neuron’s state spikes, literally referring to its electrical output. The purpose of an SNN model is to draw inferences from these spikes – to see if an image or a data pattern triggers a memory.
The spiking neurons build connections between each other as they process information, leading to better performance over time. The major takeaway here is that a neuromorphic computer could potentially handle intensive AI applications like facial or pattern recognition with a lot less power and minimal latency because the system would be able to learn without needing to reference a supercomputer-worth of CPUs, GPUs, and memory. In a sense, it would have its own kind of human intuition in a silicon chip.
Who Is Working on Neuromorphic Computing
Intel is one of the companies leading the way in neuromorphic computing with its Loihi processor that features an SNN architecture with more than eight million neurons that can reportedly perform 1,000 times better at certain computing tasks with 10,000 times the energy efficiency over conventional CPU systems. IBM (IBM) is also in the mix among the major tech companies that are working on neuromorphic chip technology, while Qualcomm has apparently backed off from the hardware side of neuromorphic computing. There was some speculation that Graphcore, an AI chip company that is part of the unicorn club of most valuable startups, was designing a neuromorphic chip but it has denied that rumor.
Neuromorphic Computing Startups
However, another AI chip startup that we’ve covered before, Paris-based Prophesee, is working on computer vision technology using neuromorphic computing to help machines mimic the human eye and brain. The bio-inspired tech can process 100,000 video frames per second for applications in robotics, industrial automation, and self-driving cars.
It didn’t take us long to find a couple more startups operating in the space.
Founded in 2012, Aspinity out of the other Virginia has raised $3.5 million in disclosed funding. The startup rolled out its ultra-low power analog processor RAMP, for Reconfigurable Analog Modular Processor, earlier this year. The neuromorphic hardware, which incorporates machine learning, is designed with the Internet of Things in mind, where any opportunity to reduce power demands for always-on sensors is generally a good thing. RAMP can detect and identify voices, alarms or even changes in vibration frequency or magnitude from background noise before the data are digitized. The ability to tackle raw data without digitization reduces the power required by up to 10 times and the volume of data handled by up to 100 times, according to Aspinity.
Founded in 2017, aiCTX is a Swiss spin-off from the University of Zurich that has raised $2.8 million in funding, including $1.5 million from Chinese tech giant Baidu, to develop its line of DYnamic Neuromorphic Asynchronous Processors (DYNAPs). Like Aspinity, aiCTX is targeting the IoT industry, especially with applications for real-time vision processing, such as gesture recognition, face or object detection, location, tracking, and surveillance. Other areas where its neuromorphic processors could be used include always-on health monitoring and auditory processing.
Another Swiss startup with academic ties to the University of Zurich is iniVation, which was founded in 2015. The company has developed what it calls the Dynamic Vision Sensor based on neuromorphic computing technology, meaning that it needs a lot less data to process images compared to conventional cameras because it only cares about pixel-level changes in real time. The result, the company says, is a “stream of events at microsecond time resolution, equivalent to conventional vision sensors running at thousands of frames per second – but with far less data.” Here’s a short demo of gesture tracking and recognition using a pair of dynamic vision sensors.
That reduces the amount of power, data storage and computing power required to use the sensor, which can be used for applications in IoT, self-driving vehicles and industrial uses such as high-speed production inspection.
Publicly Traded Neuromorphic Computing Company
We even found a pure play in neuromorphic computing during the course of our research.
Founded way back in 2006, BrainChip is the brainchild of a guy named Peter van der Made, who invented one of the earliest high-resolution color graphics accelerator chips for the IBM personal computer back in the day. Publicly traded on the Australian Stock Exchange with a market cap of about $31 million USD, BrainChip (BRN:AU) introduced a neuromorphic chip called Akida last year that implements SNN for edge computing applications in image recognition, as well as financial analysis and cybersecurity. Each Akida chip has the equivalent of 1.2 million neurons and 10 billion synapses, representing 100 times better efficiency than neuromorphic test chips from Intel and IBM, according to BrainChip. (It’s unclear if that’s still the case with the latest iteration of Intel’s Loihi neuromorphic processor.)
In 2016, the company acquired a French AI computer vision company called SpikeNet Technologies for about $1.5 million. BrainChip leveraged SpikeNet’s SNN software suite to create BrainChip Studio, which helps law enforcement to rapidly search vast amounts of video footage and identify patterns or faces. The SNN technology makes the software particularly effective at scrolling through low-resolution video to detect and classify faces. The technology is now available to license and hopefully, some deals start getting inked soon because the share price is falling off a cliff while the company bleeds cash every quarter.
Neuromorphic computing is a rapidly emerging field in artificial intelligence that focuses on the difficult problem of designing hardware that mimics the human brain in order to support the complex computations required to find cute kitten videos on the internet. Seriously, though, it’s interesting to watch the convergence of man and machine, whether it’s in designing brains on silicon or developing computer-brain interfaces, as we barrel toward the great Singularity. Progress is certainly being made on numerous fronts, but our own ignorance about how that grey bowl of jelly in our skulls actually jiggles and jives in the real world means that any effort to copy it digitally will remain a crude approximation for now.