Neuromorphic Computing: Redefining the ‘brain’ of the computer

If you ask someone, “What makes a human brain so powerful?” Most of us would vaguely describe its billions of neurons, firing electrical impulses, converting them to process thoughts, movements, and emotions. And they’d be right — the brain’s architecture is an unrivalled marvel of mother nature. Now, picture running a marathon while simultaneously playing a game of chess or solving a complicated maths equation whilst talking to 10 different people without dropping focus. Sounds impossible, doesn’t it? Yet, this is precisely the challenge neuromorphic computing aims to tackle, revolutionising how machines learn, adjust, and process information. 

What is neuromorphic computing?

In modern-day computers, the CPU is based on the well-known Von Neumann architecture, a robust and reliable computing paradigm. It splits memory and logical operations, ensuring continuous and accurate execution via the instruction cycle. Yet, as technology progresses and hardware evolves, the reduced rate of data transfer between the CPU and memory can lead to a significant performance constraint. This issue is known as the Von Neumann Bottleneck (VNB) and has been known since 1977.

To address this limitation, in the late 1980s, Carver Mead, a scientist and engineer, proposed the first instance of neuromorphic computing. Inspired by the human brain, this revolutionary approach to computing, aims to overcome the limitations of traditional systems. 

Introducing neuromorphic computing

Neuromorphic computing is a revolutionary and groundbreaking way of data processing. If you take the words apart, ‘neuro’ refers to the nervous system and ‘morphic’ defines shape and form; thus the term neuromorphic literally means “having the shape of the nerve”. In the context of computing, it characterises computing that is modelled after the structure of the brain. 

The consensus view on neuromorphic computing underlines the type of computing where processing is handled in a manner similar to the human brain. It's important to note that neuromorphic processing is not an impractical attempt to replicate the brain's architecture and function, rather, it involves extracting the known methods and structure from the brain and applying it to practical computational applications. The enactment of this type of computing has capability to transform industries, such as data processing, cybersecurity and many, many more.

The current state of neuromorphic computing

In today's world, AI is the all-new buzzword, and it feels like it's getting used in every industry, around the globe. And as the demand for these evolving systems increase, AI systems must scale, neuromorphic processing will catalyse the expansion and performance of AI exponentially. The convergence of these cutting edge technologies will allow for unprecedented possibilities and potential. With this in mind, most systems fully take advantage of this architecture and implement machine learning algorithms to a powerful and adaptable model.

For now, neuromorphic computing is at its worst stage, meaning it will only improve. In industry, there are a plethora of existing neuromorphic chips, take for example IBM’s TrueNorth chip. Containing 1 million neurons, 256 million synapses and a total of 4096 cores, this chip is a testament to what can already be achieved in 2024. The chip has the capability to react almost instantaneously to dynamic driving conditions in an autonomous vehicle, or make split-second decisions and calculations to precisely move a robotic actuator, even if it's on an unknown environment like Mars.

The term neuromorphic is more dynamic than it seems. As the CEO of Prophesee, Luca Verre, told the EE Times, “Ask 10 different people, and you’ll get 10 different answers”. It’s essential to understand that technology is developed under various constraints and ideologies thereby altering focused definitions and rationale. 

For example, neuromorphic technology is used in low-impact applications such as energy-efficient sensors. These sensors are perfect for tasks requiring quick data processing and low power consumption, because they are made to resemble biological neurons. Neuromorphic chips, for example, are being used in smart agriculture to monitor crop health and soil moisture with impressive efficiency.

Additionally, developers include features such as real-time vital sign monitoring in wearable health devices without significantly depleting battery life. These uses might not make headlines in the news, but they are prime examples of how neuromorphic computing can subtly aid in general technology, allowing for wider acceptance and more ambitious uses down the road.

How does it work?

So, how is this actually accomplished? This is achieved via artificial neurons and synapses, sending ‘blips’ or ‘spikes’ of electricity in a vast interconnected web of neural networks, which are machine learning algorithms that use data to ‘reason’ similar to the brain. These are named spiking neural networks (SNNs) and are the integral base of the neuromorphic architecture. The strength of the input decides the rate at which neurons fire; the pattern formed correlates to an output and is decoded. Pathways between neurons are formed through connections called synapses, where the strength of these connections can change over time. These changes, often described as adjusting the 'weights' of the connections, allowing the network to learn and adapt based on experience, allowing it to improve its responses and retain information. 

Furthermore, what sets neuromorphic systems apart from traditional computing is its ability to perform event-driven and asynchronous calculations. Unlike conventional processors that operate on 1’s and 0’s and follow sequential instructions, neuromorphic systems perform computations only when spikes occur.

Challenges and future directions

As great as neuromorphic processing sounds, it still has ways to go to become a fully adopted method of computing.

The difficulty of creating and producing large-scale neuromorphic systems is a significant obstacle. Scalability and integration with traditional computing architectures are limited by current hardware limitations. To add, developers face difficulties due to the lack of standardised software tools and sophisticated algorithms. Specialised event-based data is frequently needed by neuromorphic systems, but obtaining, processing and actually using this data can be difficult. In order to overcome these constraints and promote a more developed ecosystem, more research and development is a must, and is required to fully realise the potential of neuromorphic computing.

Neuromorphic computing has the potential to completely change the landscape of technology as we know it. As Albert Einstein once declared - “Computers are incredibly fast, accurate, and stupid. Human beings are incredibly slow, inaccurate, and brilliant. Together they are powerful beyond imagination.”. By harnessing the brilliance of human ingenuity and the wonders of computing, we can solve problems that we thought once would be impossible.

References

Neuromorphic computing (2024) Wikipedia. Available at: https://en.wikipedia.org/wiki/Neuromorphic_computing (Accessed: 19 November 2024). 

Ottati, F. (2023) TrueNorth: A deep dive into IBM’s neuromorphic chip design, Open 

Neuromorphic. Available at: https://open-neuromorphic.org/blog/truenorth-deep-dive-ibm-neuromorphic-chip-design/ (Accessed: 19 November 2024). 

Ward-Foxton, S. (2022) What does ‘neuromorphic’ mean today?, EE Times. Available at: https://www.eetimes.com/what-does-neuromorphic-mean-today/ (Accessed: 19 November 2024). 

(2024) | comparison of the von neumann architecture with the neuromorphic... | download scientific diagram. Available at: https://www.researchgate.net/figure/Comparison-of-the-von-Neumann-architecture-with-the-neuromorphic-architecture-These_fig1_358255092 (Accessed: 19 November 2024). 

https://newsroom.intel.com/press-kits/neuromorphic-computing/#gs.ns3z2e