Earlier this year, a team of MIT researchers revealed an exciting breakthrough in the field of neuromorphic computing––a design for a neuromorphic chip.
Scott Tan, Jeehwan Kim, and Shinhyun Choi have developed a revolutionary chip design that pushes the envelope of neuromorphic engineering, a field of study in which human biology serves as an inspiration for artificial neural network creation.
Their new “brain-on-a-chip” has the artificial intelligence community buzzing as it moves us one step closer to a future with portable, compact, and energy-efficient supercomputing, a prerequisite for many next-gen technologies, including autonomous driving vehicles and humanoid robots.
What is a Neuromorphic Chip?
A neuromorphic chip is a processing unit that functions very similarly to one of the world’s most powerful computers – the human brain.
These chips contain synapses, or small gaps, over which electrical signals of varying intensity are transmitted. This is a dramatic improvement over status quo hardware which relies on binary, on/off switch information transfer. While this does work, it is a far less efficient way of processing information.
The Verge made an eloquent comparison between traditional hardware, like CPUs and GPUs, and neuromorphic chips, which likens the former to morse code and the latter to speech. Both are effective forms of communication, but speech packs a far more productive punch. Neuromorphic chips have the added dimensions of signal intensity and analog firing, clear advantages over the fixed-rate electrical pulsing that computers today use to process data.
The new neuromorphic chip design takes after the human brain which also contains synapses, over 100 trillion of them, where electrical and chemical information is transferred between neurons. The more two neurons interact, the stronger their connection grows. This results in the formation of neural networks – literal brain rewiring – which is how we learn new information and pick up on unfamiliar patterns.
When this rewiring happens in artificial neural networks, we get machine learning.
Neuromorphic engineering isn’t a new concept, but the MIT team took a unique approach to their chip’s design. Tan, Kim, and Choi’s chip is made of silicon germanium, a material that allows for a much higher level of precision when it comes to controlling electrical current. In the past, the challenge has been in creating uniform voltage exchange, a necessity for consistent information transfer. With this new design, the team can predict ion flow across the chip’s 25-nanometer synapse which means that the artificial neurons can establish meaningful connections with one another.
For those who want to learn more about the specifics around the revolutionary chip design, MIT News put out an incredibly in-depth report that is worth a read.
The researchers received $125k in funding from the National Science Foundation and have been developing their chip for three years. Recently, they published encouraging simulation results in the scientific journal, Nature Materials, in which they trained an artificial neural network to successfully recognize handwriting samples 95% of the time using their new chip design.
Looking ahead, the MIT team wants to replicate this success by using real handwriting samples.
So, What Does This Mean?
Today, we require supercomputers with incredible processing power to facilitate machine learning and artificial intelligence activity.
Future technology with embedded AI, like autonomous vehicles, will need to assess massive amounts of information and data points in real-time to effectively make decisions that don’t interfere inappropriately with human behavior. If humanoid robots want to enhance, not detract, from day-to-day life, they will need to be able to engage with unpredictable people and continuously changing physical environments.
Neuromorphic chips like those being created at Massachusetts Institute of Technology are the portable, compact, and energy-efficient solution that will drive these other developments forward.
Although the technology isn’t market-ready yet, investments by big-name players suggest there is high-level optimism around the potential for neuromorphic chips to positively impact future innovation. IBM and Qualcomm are two examples of companies that are working on similar artificial neural networks and chips that also mimic human biological systems.
According to Grand View Research, the neuromorphic computing market is expected to grow at a healthy 20% CAGR through 2024. CNBC also recently cited an Intersect360 report that put the market size at $4.5B in 2017.
Neuromorphic chips aren’t going to overtake computer hardware tomorrow, but we will likely see increased usage and prevalence over the next 5-10 years.
What Else Will Neuromorphic Chips Change?
Although self-driving vehicles and humanoid robots are two of the more obvious beneficiaries of neuromorphic chips, there are many other implications of this technological breakthrough.
What other innovations or spaces might be impacted in the future by this level of portable processing power?