We need Bio-inspired AI and robotics -here’s why

📚 Background

AI has probably become one of the most used acronym. So much so that I don’t even need to spell it out to you. With a market value that is predicted to reach $4.8 Trillion by 2033, it is ubiquitously present and has embedded itself into our lives.

The most popular type right now involves large language models such as ChatGPT. Yet AI is a broad term encompassing a plethora of meanings. It can be divided into two broad application categories:
🔹 software (creating an artificial virtual mind)
🔹 hardware (creating an artificial body embedded within the world, also called embodied AI).

In an ideal (or dystopian depending on how you look at it) world, we could create artificial agents which are highly intelligent (in the cognitive sense) and capable of solving a variety of complex problems. This is often referred to as artificial general intelligence (AGI).

This could involve a machine that can genuinely understand and is capable of complex cognitive tasks. Ideally this machine could also run preferably on little power. Although LLMs can give the appearance of AGI, the deeper you dig, the more you realise you are hitting impenetrable ground and are often left disappointed. While impressive, there are still fundamental limitations. But how could we fill the undeniable gaps and problems it poses? The solution could lie in nature.

🧬 Bio-inspired AI

While ChatGPT does generally follow an architecture which is slightly similar to how the brain operates (neural networks having the capacity to learn), even scientists which were instrumental at building it admit that the way it truly works represents a black box. meaning that it a complex system which we do not understand how or why it works the way it does.

Instead, scientists have been considering the design of our own bodies and brains to inform the next AI models. This is bio-inspired AI. One of the main advantages of this could be that if we are able to emulate artificial life according to our biological design. This could allow us to understand how the AI works instead of it being a black box. This transparency would empower us to be more in control because it would potentially allow us to understand how its mechanism works - meaning that we can tweak its mechanism if needed.

But beyond that, using biology to inspire the design of AI could in fact create more energy efficient and cost-effective AI models. When LLMs were first introduced, the excitement gained more attention by the public than its main disadvantage – the fact it is power hungry. Broadly speaking, while the brain is capable of running on approximately 20W of power, a conventional LLM can use 3 million times more than that for just training. This should be enough of an incentive for an attempt to mimic the brain’s architecture and processing approaches. Assuming, of course, we don’t want to be at risk of a major energy crisis. Experts recognise this need for change and with a predicted market value of $13.25 Billion by 2033, bio-inspired AI could be a promising avenue for the future of AI.

🧑‍🔬 Talking to the Experts

The field of bio-inspired AI can be quite complex. Luckily at the University of Sussex where I completed my PhD, we have a vibrant community of neuroscientists who I could reach out to for real insights into the current advancements in this field, the potential advantages, and what the future might hold for it. I had insightful conversations with two talented researchers, Oluwaseyi Jesusanmi and Francesco Innocenti who are at the forefront of bio-inspired AI.

For Oluwaseyi Jesusanmi, bio-inspired AI means looking at organisms which are able to do something accurately and efficiently and then using those techniques in AI and robotics. To this end, during his PhD he has successfully built a spiking neural network (which is essentially an artificial brain designed to be more similar to how the brain works) inspired by the ant and bee mushroom body. He fed some videos and gave it a task to navigate to a specific point in a virtual environment (like in a video game, except the artificial brain did all the navigating). By emulating the brains of insects which are incredible experts at navigation, he designed a network that is not only able to navigate a natural environment but is biologically plausible meaning that it is similar to a biological design and can thus be understood more easily. In his recent paper, he shows that this was then also embedded into a robot which was able to navigate based on this bio-inspired model. This was more computationally efficient and thus would require less energy to run. As described by the researcher, bio-inspired AI can also be linked to biomimicry. We can see that biomimetic designs have yielded powerful machines in the past (such as airplanes from the inspiration of birds). While he believes bio-inspired AI still represents a niche field for now, he hopes in the future that these brain-inspired approaches will be more prevalent as they have the potentially to be substantially more energy efficient and could also help accelerate the development of embodied AI.

Francesco Innocenti’s PhD work and most recent paper explores a brain-inspired way of training neural networks called predictive coding (PC). Going back to the early days of neuroscience, the basic idea behind PC is that the brain is constantly trying to predict what’s going to happen next, and so the brain's job becomes that of minimising the errors of its predictions. Compared to the standard recipe for training neural nets, PC is much closer to how the brain might learn and could be potentially more energy efficient. However, the problem with PC (and other brain-inspired alternatives) has been that it has not been scalable. It works on small networks, but not on big ones like the kind used for modern AI applications. Remarkably, through a detailed analysis, they were able for the first time to train much bigger networks with PC, paving the way for more scalable solutions.

While Francesco believes that bio-inspired AI has several advantages, with energy efficiency being again the most promising one, he is less optimistic about the future of the field. In the early days of AI, where scientists arguably drew more inspiration from neuroscience (such as the development of convolutional neural networks), this link has been quite loose in recent decades, with AI relying more on engineering and mathematics. For example, the “transformer architecture”, which is the core technology behind LLMs, was introduced with no particular biological inspiration. Apart from the obvious energy costs, he doesn’t believe that this is necessarily a bad thing. Current LLMs, for example, know much more than any human being while having orders of magnitude fewer parameters (or connections) than brains, although in other respects they still remain much less capable than humans.

🦾 In Terms of Hardware

when we think about robotics, sensory-motor interactions are still not good enough in robots. So, for embodied AI, using bio-inspired (also called neuromorphic) hardware could be more feasible. Indeed, bio-inspired hardware can be incredibly powerful and has the potential for market wide applications. One example of such hardware are event-based cameras which are inspired by the retina. Combining these with neuromorphic attention algorithms, yields powerful sub-second object detection systems. These could be harnessed in systems that rely on very fast reactions, such as in cars.

💬 My Take

I believe bio-inspired AI has the potential to become a ubiquitous part of our lives and could represent a more sustainable avenue for AI, especially embodied AI. However, there are some limitations of this. It’s all well and good to use biology as a blueprint but the problem is we don’t yet know how many biological systems work entirely, especially brains, which are often the most popular system that scientists try to mimic. This is why I believe we need a lot more research in fundamental biology and neuroscience to be able to create reliable and performing bio-inspired AI.

Thanks for reading and until the next time,

Maria Cozan