Welcome to today’s meeting of the Science, Innovation and Technology Committee. As the leaders of the G7 meet with the CEOs of the world’s most powerful AI companies, I am sure energy consumption will be on the agenda, so it is really timely that this is the first session of our new inquiry looking at the future of compute. We will look at power and capacity, and specifically at whether emerging innovations in low‑energy computing can support a more sustainable use of artificial intelligence and machine learning tools. I want to say that we launched this inquiry following pitches to our “Under the microscope” initiative. We received over 30 pieces of written evidence, so a big thank you to everyone who took the time to contribute. Today we will be looking at the science underpinning advances in neuromorphic computing and silicon photonics in particular. We are particularly interested to hear how well positioned the UK is to ensure research breakthroughs can be translated into real-world applications. A big welcome to the panel. I am going to ask each of you in turn to introduce yourself and your area of research, and to define or explain for us what neuromorphic computing is and, particularly for me actually, how it differs from neural networks.

Good morning. Thank you very much. I am glad to be here. I am representing the academic community. I am with the University of York with the school of physics, engineering and technology. In this context I am focusing on my own research areas, which are around bio-inspired and neuromorphic computing, in particular with applications in what we call on-sensor computing. I am interested in autonomous systems, online learning, self-learning and fault tolerance, with an underpinning of microelectronics as a tool to build systems.

Did you say “off centre”?

I said “on sensor”. I will define this in a minute. You asked me to, in a nutshell, explain what neuromorphic computing is and how it differs from neural networks. Neuromorphic computing, for me at least, is in some ways what it says on the tin. It is brain-inspired computing, and in that sense it is of course related to neural networks. Neural networks have evolved over the years, with neuroscience and all the insights that we have gained from actually looking at the human brain on the one side—for example, about how the brain perceives the world, makes sense of the world and, ultimately, computes. The brain is event-based, and it is a very densely connected, complex network. In order to build systems that can compute—this is how our current AI models originated—we have simplified neuroscience and built neural networks that are in some ways simpler than the brain, especially in terms of the connectivity and topology. Everything that currently runs or is labelled as AI is typically based on very simplified networks, but scaled up on digital systems to process large amounts of data. On this journey to build these networks in this particular way, we have in some ways locked ourselves into a local optimum with this, especially when it comes to energy efficiency, because of the way the original neuroscientific ideas have been translated and have resulted in the current AI models. I can go into details, but I will let you ask.

Thank you. That is very helpful.

I am Caterina Doglioni. I am a chair of particle physics at the University of Manchester. The ultimate goal of my research is not necessarily about neuromorphic computing or photonics, but rather to understand what most of the matter in the universe is made of. Eighty-five per cent of the universe is dark matter, so we cannot see it with our current instruments. We know that it is there because the space instruments tell us that it is. My experiment is located at CERN in Switzerland. It is an experiment located at the large hadron collider, where we try to create dark matter in the lab. The issue we have is that the LHC—the large hadron collider—delivers collisions 40 million times per second. My speciality is in the amount and complexity of computing that we need to put in to find rare events in such a vast, fast-coming amount of data. I use artificial intelligence and advanced distributed computing, including hybrid computing architectures made of custom commercial chips, such as CPUs and GPUs, so similar things to what we are talking about today, but not yet neuromorphic. This is an interesting one, because via this inquiry I have learned more, so maybe we can try things. There is someone in our community who is. The fact that we are talking about this is already helping us to make connections in interdisciplinary research. My expertise is research and training of co-ordinated European training networks, and I am in a CDT, but my research is also rooted in social responsibility. That means that I want to minimise the risks and negative outcomes of the research we are doing and amplify the positives. This is why I am here, I think, because this is the environmental sustainability side of things. Social responsibility for research means making sure that the technologies that we use to answer the big questions can be used for societal applications, but also that we keep the environmental sustainability of computing at the forefront of our research. That is why I co-lead the sustainability forum for the worldwide LHC computing grid. This is the distributed computing system that we use for our computation. I am a strong supporter of interdisciplinary research. The worldwide LHC computing grid is what is funded as GridPP in the UK. This is a part of this big computing grid. We have a specific effort to reduce the impact of our computation globally, in terms of environmental sustainability.

To follow up on that, what effect might the projected usage trends of AI-related energy demand have on the UK’s energy needs and infrastructure in particular?

There is a strong energy need—there are projections that it will grow between two and fivefold. This is a global projection by the IEA.

Is that between two and fivefold for AI?

This is for the data centres that are now powering AI. There are a lot of different estimates that one can use here. The question is not necessarily how much we need now; it is whether it grows as much as it has been growing in the past. The projection says that it can grow as much and even more quickly than it has in the past. It is not necessarily just the current situation; it is the future that we are concerned about.

Are the estimates for it to be two to five times the existing data centre energy usage or the existing global usage?

This is the existing data centre usage. I need to pull up my numbers, because this is not something that I have in my memory.

That is fine. I am glad that you asked this question, because it is confusing. The whole landscape is extremely complex. Personally, I would not even necessarily throw quantum computing and neural networks into the same bucket right away. It would probably take too long to map out the whole landscape. What you said about the promises of quantum computing is that a certain type of computation can be done very fast, and only with this kind of computing paradigm and the hardware that supports it, ultimately for cracking codes, I guess, or secure communication. It is important to make a distinction between the hardware substrates and platforms running AI models and, on the other hand, the actual AI models that we use to solve specific tasks. When most people talk about AI—when we use this word, which is kind of a buzzword—what we really come to understand by that is models, which are called large language models and so-called transformer-based models. They are non-linear, feed-forward models that allow us to generate funky GIFs or summarise our texts. The large language models are very relatable to us because they are capable of producing intelligible sentences. For me, the real opportunity of neuromorphic hardware, in terms of energy efficiency, sustainability and applications, is not in replacing this seamlessly. In my opinion, and I hope I will not stand corrected now, we will not have data centres tomorrow that are made out of a neuromorphic substrate and do exactly the same type of AI that we are currently doing with our GPT models and LLMs. The real opportunity of neuromorphic computing, in its entire variety—so the different platforms, photonics and quantum to an extent, but also memristors, MEMS and sensors—is to gradually enhance and enable new AI applications. Mainly, in my case, I think these are technologies that are at the edge, or autonomous systems.

That is right.

Yes, absolutely.

There are three fundamental paradigms of what I call brain-inspired neuromorphic computing. The first is that there is no longer a separation between memory and compute, for example. In all our traditional digital systems that the world is based on, every data centre is running a digital architecture where memory and compute are usually separated in what we call von Neumann-type architecture. The second paradigm is event-based computing. In broad brushstrokes, the idea there is that, if you compute only when you are prompted to compute, you save energy. That is, again, a stark contrast to digital systems, which are driven by a clock. Even if the state does not change, you are expending energy. The third paradigm is what I call plasticity. In more general terms, it is the ability of a substrate or system to adapt and learn by changing its state. We have technologies—you can, of course, reconfigure them—such as memristors, MEMS or sensors that can actually be affected by the environment and can change.

It is then going to be physically different, yes. For example, with phase-change materials, you can control them and they will change their physical state.

T-1000. I am showing my age now.

It is great that you are taking a kind of sci-fi view on this, but it is effectively that, yes. This is the promise—far out.

Just to clarify, the materials are not actually liquid.

I do not want to lean too far out of the window and claim we have a morphable computer, but there are materials that can change state. For example, memristors change their state physically, depending on how you control them. We call it phase change. They reconfigure at the atomistic level. There are other materials as well—for example, there are sensors that rely on types of liquids. It is not as exciting as a Terminator, but the liquid is affected by the environment and how we control the device, and that is then part of a useful application.

Thank you very much. That is really helpful. As we are at this point of defining things, did you say that quantum was something entirely separate, because it is applicable only to certain types of calculation? I understand that you are going to have to semi-generalise, but is that true?

Yes and no. There is a theory of quantum gates and it has been shown that you can build any algorithm with these types of models, so a quantum computer can perform, technically, any computation. In my personal view—and I have to stress that I am not an expert in quantum computing—building quantum computers at the moment is expensive and large scale. It still requires low temperatures at the moment; there are efforts to bring this to room temperature. If we are talking about energy efficiency here, I would not build a quantum computer to build a contemporary AI model. That would be crazy.

That is very interesting. Maybe the quantum fans will get in touch with us to give other evidence.

I am sure that they will.

I would like to go back to Professor Doglioni with the figures that you were looking up on energy consumption and AI.

IT has always been something that you plug into the wall and then you feel like there is electricity coming—you need it to be powered up. But to put things into scale, data centre consumption could reach 15% of global electricity consumption. This is from an energy and AI report by the IEA—the energy agency. The thing is that it is doubling, so what happens if there is an increase in that? The error bars are very large on this, so we are not secure in the global estimates on this.

What is it now? What is the AI energy consumption now?

I could not tell you a figure because it is very difficult to find very strong methodological assumptions that hold for everything. There are also transparency issues. The companies might not want to disclose this data, because it is a competitive edge with respect to others. Epoch AI is a website that collects all the public information. The cumulative power capacity for global AI power capacity could be about 30 GW and it is comparable to the peak usage in New York state. This is quite large compared with humans.

That is a future prediction.

No, this is a projection for 2025-26.

That is what it is now—equivalent to New York state. Oh my gosh.

This is peak usage. It is not continuous use. This is something that requires a lot of energy instantaneously as well. We are talking about AI as a global problem. It is an individual problem that becomes a global problem in terms of energy. But maybe instead of “problem”, we can call it a need as well, because people need these tools. We want to use these tools. It is not the individual; it is the aggregate that goes up. There are a lot of requests for these models to be queried. Every time the model is queried, it goes to a data centre and the data centre does something in response. It is not only the model that works; it is also all the connectivity, systems, infrastructure and networks. Everything contributes to this, so it is not just the pure chip energy. It is all the infrastructure on top of that. For that, in the evidence that we sent in with Professor Gallego Schmid, we advocate a more holistic estimate of what the energy use is and what the environmental impact is. These are two different things. One is an operational question and the other is more a question of, “How are we doing in terms of net zero?” I would separate them a little. There are different questions, different problems and different challenges.

I will come back on that, but I am looking at figures now that say that the data centre capacity in the UK currently is 1.5 GW, which could rise to between 3 GW and 6 GW by 2030. Interestingly, the London City Hall estimate of power—admittedly this was five years ago—is around 6 GW as well. We are looking at significant usage now and doubling or quadrupling that by 2030. There was also a Government estimate of the amount of UK greenhouse gas emissions from supplying and constructing data centres. Their first assessment was 0.142 million metric tonnes of CO2. Then they reassessed it as being 123 million metric tonnes, so it went up a thousandfold.

There was a methodology question there.

Do you know why it went up a thousandfold?

I can pull up my notes so that I do not say something libellous. It was something that was researched quite extensively by an organisation that is concerned with the environmental impact of compute and AI. Give me one second—if you want to ask Professor Trefzer a separate question, I have a set of notes prepared for the previous time we came to the House about this.

Do you have a view, Professor Trefzer, on the impact of such a rise in emissions and why?

Are you asking why they changed the number in such a significant manner?

Yes.

I have to admit that I can only guess.

There is an exact reason. It was just how the system boundaries were defined, I believe. What do you include in this? Do you include only the chips? Do you include what goes on in the whole data centre? Do you include everything and the interconnectivity around it? Do you include only the office computers and not your GPUs? It was very much a matter of methodology and that was somehow easy to spot and revise. To me, that pointed to a question of who is checking these things. Who is checking these estimates? The estimate was made by an external company and then published. There was one internal iteration. If I or some of my colleagues could spot it, it probably was not difficult to spot. Having that layer of scrutiny when these estimates come out, and maybe involving academics more, could prevent a loss of trust in these kinds of estimates, which sometimes makes the problem more acute than it might be.

I am certainly getting the impression that there is not a consistent approach to estimating either current or future AI energy demand and that that is a barrier to taking effective action. Would you agree?

I agree.

Yes, I agree as well.

At the University of Manchester, we are working with Microsoft to iron out what our energy consumption from the use of AI tools is. It is a conversation that you need to have; it is not that you do something and then you just get the number.

You would hope that it would be something that perhaps AI could help with.

You can ask AI what its energy consumption is and it will give you a number, but the methodology is crucial. There is an entire field of study—the life cycle assessment—which is something that industry does normally.

Finally from me on this, Professor Trefzer, you described how neuromorphic networks work and the architecture. What lessons are you learning from the human brain and how it works that you are applying to the computing architecture?

In some ways, the lessons learned from computational neuroscience, or from actually observing the human brain, are a paradigm shift in the computing models and the hardware necessary to build systems that can execute those things effectively. To give you an example, which is also related to the previous question of assessing energy consumption, we can consider data movement. That is probably one of the fundamental things we could learn from the brain, because we do not have a memory bank in one half and a computer in the other. It is all one system. That, of course, has challenges, but it underpins the efficiency. The same problem exists when we talk about data movement, not within a machine but maybe across sensors and a data centre, or across our smart devices, such as mobile phones, and a data centre. There is real opportunity there. If we measure energy wrong—this is what I wanted to add to the previous question—we may miss opportunities, so this can be quite impactful, actually. Imagine if you have a neuromorphic system that is capable—on a smaller device, independent of a data centre—of performing some of the intelligent functions that we currently rely on a data centre running somewhere remotely, with a large language model or a transformer model, to solve. We have an energy saving right there. There are other benefits. You do not have to move data across communication networks, so that saves energy as well. There are opportunities for increased privacy and personalisation. Another paradigm that we have learned from the brain is that there is really a substrate—our brain—that is not a rigid computer clogged up by a digital system. This motivates us to build computing systems that are adaptable in additional ways, other than just by uploading another program, for example. This allows us to make those systems more robust and potentially adapt them to be more efficient in certain circumstances. It very much depends on applications. As I said before, it is not mimicking the transformer that runs in a data centre; it is more about enabling new applications. For example, if you want a safety system for a building, you want an early warning system that is based on sensors dotted around the building. Current solutions have the sensors all there, then the data is transmitted to a data centre, then we run massive models. The paradigm shift would be to have sensors performing autonomously. If one sensor screams in the corner, that is enough to know that we have to evacuate or that something is wrong.

I have one quick question—actually, I am never quick, but I will try to keep it quick. I always worry about analogies with the human brain in computing, and I have grumped about neural networks for years. The explanation that you gave to Kit of the substrate or the hardware is, I would suggest, more like how an individual neuron works. You have internal processing within a neuron about the incoming signal, and then it will make a decision on whether to fire or not, so it can grow out its synapses and dendrites as and how it wants to do. For me, the analogy that you gave with the plasticity, the processing of memory and actually processing in one place fits more with the neuron. I would guard against giving the impression that the brain is an amorphous thing where memory and processing occur at once, because there are different places. Does that understanding of your explanation help to fit a bit better with my understanding as a neurobiologist? I would also point out that, there are theories that hyper neurons and quantum processing occurs in the brain, and other parts of the body have neurons that can do some of the processing. At the neuronal level, I can see how the analogy works, but I do not know what your thoughts are on that or whether it is worth you exploring the internal processing within the neuron. You might get a better analogy.

Thank you for this question. I have to now be careful not to get very excited and dive into a scientific discussion. You are quite right: there are many mechanisms in the human brain that are not yet understood. One thing that I would take a slightly different view on is the importance of solely the neuron in the brain as the main contributor to any high-level computational function. It has been shown that the connectivity in the brain and the delays that this causes in the neuron communicating has a significant role. There are additional mechanisms—this is basically neuroscience—that bring in hormones and whatnot that modulate or temper brain function and put bunches of neurons into a high-alert state. The way I want to respond to this question is that this indicates to me the importance of fundamental research alongside all the translational activities that we are talking about here. Every analogy is always trying to explain this and link these up. Sometimes the analogy works and sometimes it does not. It is hard to pick one that works for everybody because we have not fully understood everything about the brain. I would argue that the brain proves its efficiency and usefulness just by its existence in 10 billion—I do not know—human beings who work more or less well. I would argue that the main goal is not to translate this one to one, or even necessarily give up before we have fully understood it, but to take inspiration from this along the way. If we find useful mechanisms or paradigms, we explore the different axes and how to translate those.

I hear what you are saying. I think that “neuromorphic” is probably just a sexy term that is quite useful. I will come back to it later on silicon photonics.

If I may correct myself from earlier, I have now found the article and I believe that, in a different inquiry, you will have the person to talk to about this. The reason why there was a discrepancy in the estimate was not necessarily due to how much was included in the estimation but how the electricity was translated into greenhouse gas emissions. It was assuming a very decarbonised grid and no gas. At the moment, the issue is that there is not enough connection to the grid, because that is already being used, so people use more gas. They use gas turbines, especially in the US. You can also go very much more pessimistic than what the Government have right now. It depends on the access to the grid.

Thank you. That is very important to know. It is not a matter of what was included in the estimates.

No, it was the translation.

It is how they are making over-optimistic assumptions about proportion of green energy that was used. That is very helpful.

I wanted to come back to the point about personalisation and neuromorphic computing and get you to talk about that a little more. I can understand that, in an academic field, it is almost like you would input all the different bits of research that you want and have accessibility to that—it becomes your own little LLM, really. How would that work on an individual level beyond the academic environment?

Everything that we talked about so far—the complexity of the hardware substrates and the models—is always pitched against a very mature technology like LLMs, which are running in data centres, but suffering from all the energy and sustainability problems. There is this tension, but the only way to have an impact and to prove its worth in the short term is through specific applications. The personalisation, for example, could come in a wearable device such as a hearing aid. Currently, these devices are built on digital substrates and we train models offline, but you could imagine a neuromorphic substrate—I disagree slightly on whether the term is just sexy, but there is inspiration—that is, for example, susceptible to sound but, at the same time, has modalities or behaviours that allow it to function in a more nature or brain-inspired computational manner. Then you could push functionality out of the digital system into, in this case, a sensor. There is significant potential that that would be much more energy efficient—potentially by orders of magnitude. At least parts of the system could then be more tightly integrated. I am giving an example and of course you can say, “This is one example.” Because of the complexity, the short-term impact and the benefits will come where people take inspiration from the lab, identify a use case with a hybrid system or hybrid integration, and marry that up with current technology to optimise the benefits. Once we have learned enough, hopefully we can do the academic thing again and abstract it.

Taking your example of a hearing aid, it could learn from you what your reaction is to different sounds, or where you live and what kind of city or rural environment you live in, and therefore be able to interpret sounds more accurately, because it knows your surroundings and your reactions to those sounds.

That would be the idea, yes.

That is as opposed to a huge model that could make mistakes because it does not know that you live in—I do not know—Runcorn.

Or Milton Keynes. Why not?

I have a second question there, because this is quite important for the Committee to understand. This is such cutting-edge work that you are doing. How is the research funding responding to these new areas? Are they within the 10-year funding commitments? Is there the right funding environment? Do you find that UKRI’s reaction is supportive enough for how quickly you want to develop this field?

This is an important question. I can only answer it from a personal viewpoint. I have this in my notes actually. The current landscape is looking promising for neuromorphic and low-energy computing in this manner that we have discussed. We have the innovation and knowledge centre, which has been funded. We have the neuromorphic hub, on which we will hear more details later from my colleagues. We have the NeuMat network, which is run out of Cambridge. There is a focus on silicon photonics and memristors from my point of view. This is really promising and builds on the UK’s strength with the neuroscience-driven research that led to SpiNNaker and things like that. There is a lot of knowledge on how to do what I evangelised about earlier about taking inspiration from biology. There are things where UKRI could do more. There is a current gap that I perceive. Since SpiNNaker is now there and we have invested more into the lower level of the stack with neuromorphic, we have slightly lost the link and the opportunities to take innovations out of the lab and across the whole stack into applications. What more could UKRI do? As I said, it would be to try to fund across the stack in a way that enables people to interact with industries. At the moment, spin-outs, for example, are a good venture to try an idea in a specific application area, perhaps. The culture in spin-outs currently is that they are well supported by universities, but the aim is to get VC funding, develop an exit strategy and sell the company. This is not going to give us a second Arm or something like that, so the strategy is not building capacity long-term. Finally, you talked about 10-year grants. It would be nice to have longer research programmes. One thing that I would put into the mix to consider is to make sure that these programmes are shaped in a sufficiently flexible way, because halfway through I am sure that research outcomes will suggest that the goals have shifted. There is also a balance to be had between inclusivity and exploitation. Which stakeholders are involved? How equitable is access? How open is it?

Professor Doglioni, I think you are a particle physicist by training. Professor Trefzer, what is your background?

I am a physicist by training, and then I became an engineer and a computer scientist. It is all a bit blurry.

We have had discussions and concern raised, particularly by the particle physicists and others in the PPAN community, as it is called, about cuts to research funding and opportunities, particularly for early‑career researchers. Is this something that will impact neuromorphic computing?

For sure, yes, because you will not have people who want to do R&D. This is one of the issues—that one person just keeps the lights on on these machines and big investments, and then someone else is doing the cutting-edge research and going in the future direction. This is a very big issue. The other thing that I wanted to say about that is, alongside fundamental research on the materials, the commercialisation and everything that Martin said so well, there is a place for interdisciplinary research to ask big questions using the technologies that others can develop. Industry will not ask the questions that say, particle physics or the Human Brain Project will—something that is really big and wants to understand something deeply—and then you really need to push the technology. After you have pushed the technology, you can take it home and say, “Now I will think about what I can do with this”. This is the case with a lot of medical technology. This is the case with the web and all these things coming out of particle physics, but they are not just particle physics. There is fundamental research that ends up being put in a bucket where it cannot be supported, because it is only the methods and the things that would lead to something; it is not the question that would push the technology.

The Committee remains concerned about the funding future for particle physics, astronomy and nuclear, and we will raise this with the Secretary of State for Science.

Good morning. Like others, I am struggling to get my head around these concepts, but they are very exciting. The issue that the Committee has been addressing is the energy use of other systems. My question is, how far down the line is this process? Is this a theoretical concept that could happen in a few years’ time? Is it something that will be adopted fairly quickly? What are the barriers to it being adopted to tackle those big energy problems?

Are you asking in the context of data centres or specific applications?

It is both, to some extent. It would seem to me that there is also a question around the major companies that have invested huge amounts in the current approach. Is it an opportunity for them or a threat? If it is a threat, what impact does that have on the ability to progress it?

That is an excellent question and of course a very complex one as well. Whether this is a threat or an opportunity depends on which company you are. As you rightly point out, if I was a big, leading GPU company, I would want to sell as many of those as possible. Coming back to where the real opportunities of what we call neuromorphic computing might be, in the short or medium term we are not replacing the digital systems that we currently have with neuromorphic ones. However, there is evidence in models from the fundamental research side that some of the neuromorphic paradigms, for example, can work alongside our current transformer models and make the learning process more efficient. That can make the transformers more robust, in the sense of hallucination or accuracy and how efficiently the training process works. This is my personal view, but, for me, these are opportunities to make short-term gains. We would have to then be open to the possibility of having an experimental data centre where we try these approaches out.

I don’t know if this is the last thing that I will be able to say here, but there is one aspect that some have brought up in written evidence, which is the rebound effect. We need to consider that when we are thinking that a technology will make things cheaper or more efficient, because people will use it more. This is something that has been proven from the industrial revolution. It is there. If you make fuel more efficient, people will drive more cars.

At this point, this is a question that is not for us—academics—potentially to answer. It is about how these methodologies and AI are used responsibly. Do we want to think about a system where we are sufficient rather than plentiful? Decoupling the natural resources needed to achieve growth from growth is going to be difficult. I do not think that there is anything in the literature that says you can achieve infinite growth and have net zero at the same time. This is very much a policy question, and maybe a global question, of where we go as a community, a population and a Government. I will leave that with you, so that it is something that enters the discussion as well.

I am hearing that you see it as complementary rather than a threat to the existing approach. Going back to the energy point, if we have the huge spike in energy demand that we are fearing, what is the connection with the work that you and others are doing to make sure that the people who are causing that huge spike in energy demand look to you as part of the solution? Is that linkage there or not?

To me, that link is there, but there is maybe not a completely satisfactory answer that says, “We will solve all the energy problems of compute.” This comes back to one of the opportunity spaces that we have discussed: specific applications that are ubiquitously used. This is not necessarily solving your data centre problem, but it will impact significantly on the energy use of certain devices or applications, which will increase as well alongside that. I am saying that it is not just the energy consumption of data centres that we need to be concerned about. It is also the amount of data produced by every single system that then needs to be transmitted, stored and processed. There are shorter-term, more tangible gains to be made. The second thing is that this is a complementary technology. Thank you for your comments, Caterina, because we cannot see this as just about thinking, “What technology can we put into this data centre to solve our energy problem?” The technologies have potential—so we can create more efficient models, potentially, to more efficiently harness the current hardware. That is, in my view, a short-term goal that could make significant inroads. In some ways, academics are happy to do it, but it will not be funded if there is no incentive in the way that Caterina just said. This is the main blocker—do not see this in isolation. Education, for me, comes in there as well because, unfortunately, many people perceive AI as something that it is not. Of course, there is also what we use AI for. It is not just the technology’s responsibility to save energy. It is also behaviour, and we can educate people. The third angle that I have in my notes is upskilling. I know that there has been movement and there was just investment. We welcome that, because having people in that skilled pipeline who can also educate others on what you use AI for and what you do not is crucial.

At one point there you said, “This is the main blocker”, as far as you were concerned. What is the main blocker?

I was referring to the fact that, if you talk about translation, the short-term gains will be in specific translations but, to make that happen, somebody needs to see a business case in it. When I said “blocker”, I meant that that business case might not always be purely economical. There needs to be a will to be sustainable, even if it is more expensive, and to be more application-specific, rather than completely generic, because that is just physics; that is just the world’s laws. We cannot have everything. We operate in a Pareto space and at the moment we are optimising for money.

That raises a key question about the role of the Government and the role of these very well-resourced other parts of the sector. On that, the Government issued their AI hardware plan recently. Do you think that that is helpful for what you are seeking to do? What is your reaction to it?

Personally, I think it is very helpful. I am not completely altruistic. In this case, I was approached by colleagues in Southampton with the SoC labs to get involved in upskilling engineers in microsystem integration, microelectronics and silicon-based platforms. For me, as an educator and researcher, the plan is key, so I welcome this investment. The pitch is very much about sovereignty and things like that, which is great, but, ultimately, it goes in the right direction and is actually putting the foundations in place.

There is a huge opportunity for also in standardising the environmental sustainability metrics around procurement. That is mentioned in passing, I think, about materials, but there is going to be a competitive round. What are we going to get people to compete on? Is it just the computation per watt, or are we going to have computation per watt based on emissions as well, depending on various scenarios? This is something that we are trying out at CERN. It is responsible procurement. It is something that I keep telling people in the digital research infrastructure about, as I am a member of the UKRI advisory board. This is something that we can test out. Maybe the first time it does not go through as one of the metrics, but it would at least get people to think about it. It is extra work, but this is a positive role that the Government could have, saying, “Let’s try this. Let’s try to make people aware of what this costs and make decisions eventually based on that.”

I want to ask a point of clarification on the timelines. You said in the short term or the immediate term, and I want to understand what you mean by short term.

This is based on experience, but this is how I define short term. Short term for me means something between two and five years. Once you are engaged and have developed an idea, then you have partners to translate this. That is what it takes to have a viable prototype. Long term is 10‑plus years.

That is an important point of clarification. We have, as you know, AI companies looking to mobilise $4 trillion, $5 trillion or $6 trillion of capital investment based on a business model that involves deploying billions and billions of pounds into data centres in an existing framework. That is looking at timelines for the next three to five years. You are saying that this technology could have an impact on the business models of AI in the next three to five years.

I believe so, yes.

There also needs to be thought about the algorithm. The way we do AI right now, or the way we think about AI, is generative AI. It is very large models that do everything for us. That is the way we are thinking about it. We could think about AI differently. We could make business models that are different or that do one thing better with a smaller model. I think that you will have a lot of evidence later about that. If we can shift that and be good at that, maybe that is part of how this shift can happen. There are the first LLMs on neuromorphic architectures, but it is unclear whether these are going to be as powerful as the biggest model from Anthropic and so on, so what are we trying to do here?

Data centres in space are a bit controversial, or a bit science fiction.

I hesitate to comment on the data centres in space. Generally, reducing the amount of communication or data that needs to be stored will translate into more efficient applications and ultimately lower power, not least because you do not have to fire up the data centre for a trivial—I call it trivial—or a simpler or more bounded task that you may as well just do with the data at hand. One thing that I want to raise for consideration as well is that we then have to accept that, just like with having goals other than economic goals, for example, the data is then not necessarily stored somewhere where somebody could go back and audit that data. We have to be open to compromises in that sense. This is where there are a lot of headaches for you, because there needs to be governance policy around this. What happens to medical data? What needs to be stored? What does not need to be stored? These are not necessarily technological questions.

No. The reason we collect data is that we want to use it to understand something and make the right decision. Arguably, once you have made that decision, you could delete the data. It does not matter. You made the right decision. If you were more ambitious you could say, “I want to use the data over time or make different decisions over time to build up some sort of model that allows me to get better at making decisions.” My point was related to where this happens. Currently, that happens with a large model in the data centre. We are collecting oodles of data and shipping it to the data centre. The data centre then compiles a large model by fitting everything. That is currently highly energy consumptive. The flip side of this would be to have a smart device that, within its scope, can produce a smaller model. It can gather data locally. It may also have the capability to build some sort of internal model that allows it over time to make better decisions. The idea would very much be not to have resources to store every piece of data that this has ever seen. We would thereby be able to have a smaller and more efficient device.

Yes. This comes back to Caterina’s point of being good enough for a specific application that allows us to make gains.

Also, at this point, everyone is hitting the same data centre with questions. If you are on a smaller device, that consumes less power, because fewer people are hitting that, but there are more devices somehow. At this point—I hate to be the person that breaks this argument—you have to think about how much it costs you and the environment to build these devices. That is something that you need to put into the mix. One can reach a break-even point and say, “I have all this research and development, and all this embodied carbon, but, ultimately, I will be doing a better job at environmental sustainability.” It just needs studies.

Quite a lot is already happening. There is a lot of AI on edge and AI on chip.

We will be looking at that in the next panel.

That is for the next panel. We need to get to the next panel.

I will have to charge consultancy now to give my opinion. All the different technologies have their different strengths. I do not want to comment on photonics because the experts are sitting behind me, but it is often the case that every technology offers specific opportunities. What we need to wise up to is how to integrate them to have the best outcome overall. I do not want to hedge my bets now.

The orchestration point in the AI hardware strategy really hit the mark. How do we orchestrate a system that has all these different technologies? This requires investment. You cannot just plug them in.

That is one question, and I think that is a global question. A question for the UK, particularly now, with all these discussions about sovereignty, is whether this is an area that the UK can lead on globally—yes or no.

To our knowledge, yes.

To our knowledge, yes.

That is fantastic. I am sorry that we have kept you longer than we intended to, but it has been such a fascinating discussion. You can tell that all members of the Committee are really interested by the opportunities that the work that you are doing raise for the UK and globally, so huge thanks for joining us today. Witnesses: Professor Sergei Turitsyn and Professor Eiman Kanjo.

Welcome to this, the second panel in today’s Science, Innovation and Technology Committee session looking at the future of compute through neuromorphic computing and photonics. I am going to ask Emily to open the session.

Can you briefly introduce yourselves and tell us a bit more about how your areas of expertise can lead to low-energy computing? This is a big issue, particularly for young people who understand or are excited by AI, but the environmental impact is something that comes up in pretty much every school I go to now.

Good morning. Can I start with some good news on the question from the previous session? You can use AI models or language models without connecting to the cloud. You can even use your wearable. Some of my students are doing that, using a technology called TinyML, which is my speciality. I am a professor in tiny machine learning at Nottingham Trent University, where we try to bring the power of AI to small devices. We call it tiny machine learning, which is almost equivalent to edge AI, or the edge of compute, let us say. Tiny machine learning is even smaller, so you could have a wearable, small device that can operate by itself without connecting to the cloud. I try to promote the idea of decentralised AI to protect privacy and also for energy efficiency, because, once you run the models on the device, you do not need to connect to a server, for example. We often talk about a “no data” type of infrastructure. Everybody talks about data sharing. There are cases where there are no data. What I mean here is that, when you are using a wearable—I do not know whether any of you have a wearable at the moment—you can, for example, use a camera. The camera will see, for example, a few people here. It will count the seven people, and then discard the image frames and give you seven people. It will be just the people count, if that makes sense. In that case, the data that we used for this model has disappeared. We do not need it. We do not need to share it. That is great news in terms of privacy because you can, for example, self-manage if you have a health condition. If you are trying to look after your lifestyle and you do not want to connect to a server, you can do that. I started my research many years ago. I had been working on mobile development and was one of the early academics who looked at using mobile phones as research tools. It was in the early days when Nokia had almost no GPS, no Bluetooth and no colour screens. It was very primitive. We celebrated, for example, when we had GPS and started to see Bluetooth. Over the years, we have made amazing progress in terms of the development of technology—in particular, microcontrollers, for example, that can run models. Mobile phones are edge devices, but they have more power and are very sophisticated. In fact, I would go as far as to say that this is almost a data centre. Some of the work we do is connecting devices to make them data centres.

That is fantastic. Professor Turitsyn, could you introduce yourself and your field?

I have two posts. First, I am director of the Aston Institute of Photonic Technologies. Photonics is a very nice science, of course, connected to quantum, as you mentioned, because quantum is actually just a few photons. It is also an important part of the UK industry already, at £18 billion, with about 70,000 people working there, and about 1,400 small companies. It is a very good contribution to UK prosperity. As far as AI is concerned, photonics is also very nicely connected. Eiman mentioned cameras. Cameras are about photonics, because they collect photons. Why are photonics and AI naturally connected? Because AI is about data, apart from small applications on the edge. The more data, the better, because you can learn from data. You can find patterns. You can do a lot of good things, as you have observed already. What science or technology can generate data at high speed? The speed of photonics is the speed of light. It is the highest speed possible. That is why photonics is the main producer of big data. It is for imaging, sensing and communication. It is already happening. Photonics is also important because it transmits this data through fibre optics everywhere around the world, in the data centres. You will have noticed recent purchases by Nvidia and other big photonics companies, or collaborations with companies such as Corning, because it is important for data movements, which have been mentioned already. Photonics is now also moving to do computing. In photonics, we have devices that are relatively complex. They are non-linear. AI can help us to see some patterns and information about these devices that we can overlook, even if we know equations. It is, basically, a perfect match between two sciences or technologies. In my second post, I am director of the recently founded UK Multidisciplinary Centre for Neuromorphic Systems and Computing—NeuroSYNC. We co-ordinate this centre at Aston University. It is a consortium of seven universities, including Oxford, Cambridge, Strathclyde, Loughborough, Queen Mary University of London and Southampton. I did not forget any—phew.

Well done.

Of course, we have a lot of industrial partners, including big ones in defence, as well as BT and Microsoft, and some societies that represent neuroscience. This is multidisciplinary. It is a really important programme implemented by UKRI. Our role is to bring together these different communities, which might be important for neuromorphic computing. On the one hand, we have neuroscientists working with stem cells or organoids to learn something about real neurons in real biological systems. We have people who develop algorithms, based on these ideas and on general understanding, with inspiration from biological systems. We have people working with electronics, with photonics, and also with different sorts of applications and, of course, materials, which are very important. Can I pre-empt a question from you about silicon photonics? Silicon is a very important material, but there are many others that might also be interesting, so we will talk about that. In this second role, our role is to create road maps for research in neuromorphic computing, because it was mentioned that there might not be an immediate application. It is also a long-term vision to create a road map for industry and policymakers such as yourselves. I believe that, in October, you will have either the first draft or an already prepared road map. Also, we have part of our group working on very specific applications, where you need some access or clearance for defence, because that is also very important.

In your role as head of these two large institutions, what is the role of ethics and sustainability in your thought processes when developing road maps and your calls for supporting particular technologies?

It is important. It is now everywhere, as you know. Of course, in our road map, we will address these issues. It is for the full life cycle. Let us take ICT as an example, because it is huge and is doing a lot of good things. After that, you have some devices that are not used, and we need to utilise them. Fortunately, at Aston, we also have the institute of bio-research and energy. We talk with it a lot at the moment about green stuff and bioenergy, but devices that are not used can potentially also be used, with appropriate technologies, to create energy.

What about the ethics side?

The ethics side is probably most relevant to AI applications. At the moment, I do not think that we have any concern about neuromorphic computing, but there is one important question, and it is not about neuromorphic computing. It is about AI in general. You mentioned that we now understand neural networks, but we do not know how these neural networks, with billions of parameters, make decisions or outputs. This is quite worrying. I am not talking about judgment day or Terminator, as you mentioned, but we need to better understand what we and our devices are doing.

On the ethics side of things, because we are working at the forefront of technology, we discover something every day. In particular, I am trying to understand the implications of current technology on cyber-physical security in particular. With the advancement of agentic AI and these devices that are autonomous and that can evolve, there is a great concern about their capabilities. This is an area of concern that we have not looked at extensively.

You say that neuromorphic technology does not raise any ethical concerns. When we heard about how hearing aids could be personalised through collecting the data of an individual human, that did raise some ethical concerns with me. I am sure that, in any transformative technology, ethical considerations should be part of the assessment or the ongoing discussion and debate. That is part of what we are here for.

I just want to focus on silicon photonics. How can advances in silicon photonics help to reduce AI-related energy demands specifically? I am thinking a bit about neuromorphic. I can see how, in the signals, instead of a nice neurotransmitter slowly diffusing across a synapse, you could whizz it back with a photonic signal and it would be really quick, but I might be oversimplifying it. How could silicon photonics reduce AI-related energy demands?

As I said, silicon photonics is just one important part of material photonics. There are other interesting materials, such as indium phosphide, silicon nitride, or thin films of lithium niobate. In terms of how photonics can help reduce energy consumption, it is already under development, including by big companies. In electronics, there are many advantages. It is working already and we know that, but it is always about some wires. You have wire and you have electrons going along the wire. In photonics, you can have huge parallelism. You can send photons in parallel, such as our images. I am sending photons now, although they are probably reflected photons coming from the lamp, but it is in parallel in space. It can be in parallel in the spectral domain, in frequencies. It can be in parallel in different situations or directions. This parallelism is key in how optical devices can accelerate the work of normal digital electronic computers without expanding the energy use hugely, because they do this work in parallel. They are called optical accelerators. This is why they are under development. When we talk about immediate impact, this would probably be the closest one.

Because you mentioned that it is not just silicon, I am going to jump to how—

Silicon is important.

My next question is looking at manufacturing and the industry, and how well the research has translated into it. You mentioned that it is about £18 billion and an incredible number of jobs—I have 84,000 in my notes. It is one of our most productive UK manufacturing industries. How well has the UK translated its research strengths in photonics, including silicon photonics, into manufacturing and industry? It seems to have done that quite well, which is quite unusual, because we do struggle with that sometimes. I am also quite interested in what the main barriers are to greater deployment. Because you are talking about these critical minerals and other types, I am quite interested in their supply chain as well. It is a bit of a broad question.

There are some really critical minerals such as rare earths, but silicon is, basically, quartz. Fibre optics is, basically, quartz. It is sand, so it is very cheap. That is why it is good for businesses. There is also a lot of sand even on our beaches around the UK, so it is a cheap mineral. In this sense, it is less affected by the criticality. There was some tension between the US and China over rare materials, but that is less relevant to silicon photonics.

So we have a different issue with things such as lithium from that with silicon photonics. Going back to my main question, one of the things that came back was a criticism that the UK suffers from not having much in the way of a prototype packaging line—that was from Professor Penty of Cambridge University, and Cornerstone has agreed with that—and that we should seek to establish a national silicon photonics pilot line, which would enable innovators to anchor high-growth ventures domestically and create a sovereign capability. What are your thoughts on that? Is that something we need?

I agree. We need prototyping. We need packaging. I would still insist that it is broader than just silicon photonics. I would say integrated photonics, because we should not exclude any other potentially important materials. It is not the role of Government to choose a winner and say, “We will go with the winner.”. It is business that should choose the winner. I am not the Government but my view on research support is that it should be relatively broad. It should support different solutions. Regarding infrastructure facilities, it is important to invest and not to stop after you invest. Again—I might be wrong, so do correct me; you are the politicians—sometimes, political action is saying, “Let’s invest £100 million in something”, whether that is fast internet, that cancer will be solved, or a big building. You take a photo, then job done, so you go on to the next question. The real question is how this facility will work, and how it will help small companies and other groups that are not part of running these facilities. When you create some unique facility, you are creating a sort of monopoly, because the people running this facility are probably tempted to put the cost of running it just on its users. There should be some follow-up procedures to deliver what it was meant for. If you say “national facilities”, that should mean that it will be accessible to small companies that cannot afford it and that is why they cannot go in certain directions, or that it will be available to some research groups that do not have clean rooms for whatever reason. The cost for them should be affordable. Why would you create a national facility that nobody, or just a very small group of people, can use? Another important consideration is that, when small companies go with their idea to a national facility, we need to ensure that their IPR is protected and that it will not be part of a joint collaboration where the IPR will go from them. That is very important. This is about a follow-up procedure. If you decide to invest, I am pretty sure that you will make right decision, but try to ensure the next steps, because how it is run is very important. You can evaluate it. If you invest in national facilities, one KPI should be how many small companies or research centres outside of the host used it.

Your points are really well made and a lesson for us all about the follow-up and the IPR, as well as scaling up, which we are not so good at.

Can I break with normal procedure? You asked a very important question about the definition of neuromorphic. That is really important, because neuromorphic is normally associated with something we learn from nature—from the brain. What we are discussing here is much broader, and there are different terms. You said it in the right way—it is just a fancy term we use, and we use it because it is there already and we have created these centres. We do not want to lose this term, but we should explain that there are different terms that are also relevant, such as unconventional computing or analogue computing. This counting was an analogue operation, not digital. Now, when I send some information to you through sound waves, it is analogue information. There is no digitalisation. Then it gets to your brain and you get this information. There is no digital involved.

Professor Turitsyn, thank you very much for that clarification. We do need to be moving on as we are running out of time. Allison’s question was specifically about the barriers to investment and to the deployment of photonics and silicon photonics. You identified barriers that we have seen before as part of our inquiries into regional research and growth.

Water consumption by current data centres is related to cooling. In this sense, photonics will be better. It will not require that level of cooling. In terms of toxicity, I would say that nothing at the moment is comparable with the consequences of this water or energy consumption. There are some materials that are toxic, but they are used in laboratories and special environments. When something goes into mass production, it is normally not toxic. It is something that can be easily used.

Again, it is quite clear. It is a material in which we can see sand, or quartz. You do some heating and some processing, but it is nothing really toxic or heavily chemical.

Photonics is an enabling technology. It is almost everywhere. Think about electronics: you are not saying electronics is just one sector of the economy. Electronics is everywhere. But also, in your phone, you have photonics—in the camera and so on. During the International Year of Light, it was well noticed that behind the photonics component in the phone, there are 26 Nobel prize winners.

You are talking about cameras and so on, but we are talking about photonic computing.

That is correct.

Because it is not yet technology that can be used immediately there. You might also notice that all these companies are buying photonic companies or collaborating with Corning or Oriole Networks. For data movement, photonics has no comparison. Data movement in data centres will be done, and is probably already done, by photonics—apart from small scale.

No, I did not say that. It was already discussed, and I completely agree that that would probably be the wrong idea. There will be new data centres where there will be optical accelerators that will do, for instance, matrix multiplication in parallel. Again, we have some companies in the UK that claim some record results. There are the first prototypes, but they can be developed into mass production, probably in several years. I do not think it will be about replacement; it will be coexistence, and for a long time. We should understand that electronics and digital were developed over many years, with a huge investment. That photonics is even considered to be at this level tells us something about the potential of photonics.

In general, we tend to focus too much on one layer of the development stack—for example, photonics that will accelerate or reduce power consumption. The technology might not be ready in a few years or, to my understanding, even in 10 or 20 years. The other issue is how you embed this in the system. If you have the best photonic or analogue chip in the world, you still have to find a way to fit it in the system so it can work. That involves data centres, for example, or a mobile phone. Besides that, the error ratio is very high if you are working on an analogue photonic device. At the moment, I am collaborating with a range of multidisciplinary researchers. I am not expert in analogue, but I am architecture-agnostic, so give me a chip that can run something, and I will take it and put it in the system. At the moment, we are using digital and simulating analogue because we cannot put our hand on an analogue device, whether it is photonic or resistive. One of the problems, as I said, is technology readiness. The second one is, even if it is ready, how you integrate it in the system and how stable this technology is. That is why, as you mentioned, it is not about this versus that, but about a hybrid approach where digital technology can enable the progress of analogue or neuromorphic technology.

The carrier is a photon. We are exchanging photons now; they are also in fibre optics. The internet is already on these carriers, which are optical pulses. Can I add something to support what was said?

I just want to clarify something that Professor Kanjo said. You talked about 20 years.

I just mentioned 20 years but, as I said, I am not an expert in photonics or analogue in particular.

Professor Turitsyn, what is your view?

One of the immediate obstacles is a lack of software compared with digital. Digital computers are not important as boxes; they are important because there are many software programs already running on them. When you have new hardware, you will also need to create an ecosystem of programs that run on this software, and I do not expect it to be immediate.

Thank you very much. Professor Turitsyn, you spoke earlier about the funding approach of the Government and not picking winners. You are saying that there are barriers to diffusion in terms of software, and that you are looking at a 20-year or long-term horizon. I think that is what you said. What could the Government do to reduce the timescale for deploying edge AI capabilities?

I did not say 20 years. It was Eiman.

I did ask you to clarify what time you thought.

First of all, DSIT, UKRI and Innovate UK are doing a good job. This is why we are here. Maybe we are a little bit behind the Netherlands, but we are learning and are a very short distance behind. What is probably missing are more interdisciplinary projects in which people with a background in AI and in using digital are working with photonics. When we better understand what the challenges are, maybe we can contribute, and they will understand what our challenges in software are. Basically, the pairing of experts and researchers from different areas would be huge. If you asked me a question about commercialisation, I could answer, but you have not asked that yet.

My question was about the timeline that Professor Kanjo suggested, which is about two decades, for the diffusion of photonic computing as a technology.

I believe it is shorter, which means that, in three to five years, it will be part of the data centres. It will be in the form of optical accelerators, which will use the parallelism of photonics to do some important operations, such as matrix multiplication, that take a lot of energy and latency. It also was mentioned already that it is not only about energy consumption, because there are applications where you need low latency or an event-based event. One example that has not yet been mentioned is the driverless car.

I did want to come to the driverless car. If this technology were to be diffused within, say, five years, do you see it replacing existing data centres or do you still see it as being in parallel with existing data centres?

I see it as being in parallel. I certainly do not see it replacing them. It will be more hybrid. It will be new data centres, but I do not think it is even technologically right to remove something and build a new one.

Professor Kanjo, talking about autonomous vehicles, which will have to make complex, important and ethically concerning calculations and decisions in a vehicle—preferably, I would think, without accessing the cloud—do you see photonic and neuromorphic computing as part of the solution to that?

In the next few years, I do not think so, but in the very far future—because, as I said, we have been trying to work on the hybrid approach, bringing together tiny machine learning and edge AI, whether it is photonics or analogue. At this time, it is not even possible to access a chip or a tool that allows us to explore whether the photonic part or the analogue part can work as an accelerator.

So not until in the long-term future. How about you, Professor Turitsyn?

In terms of time, I agree. In terms of potential, I do not, because cameras generate an optical signal, and before that goes to electronics and digital, you can do a lot of important work. That will be developed, but in terms of timescales, it is probably not in the next two years.

This is a similar question that we asked the previous panel about skills. We hear about the critical importance of physics, and there are concerns raised about the apparent cuts to physics funding, particularly for early-career researchers. What are the key skills that your area of research needs?

I could speak from my side of things. There is the perception that there is a lot of funding for AI and digital computing, but there is not. A lot of the available funding sometimes does not go directly to, for example, AI PhD students.

What skills do you need?

Hardware for AI—the system type of things, bringing things together. It is the stuff that we do in terms of tiny machine learning, so understanding the hardware, informing the hardware designers of what needs to be done, and AI optimisation. It is the whole system. For example, ChatGPT came to the world out of system development.

You need more investment in skills for systems research, AI and hardware.

Bringing the whole thing together.

Sorry to hurry you, but is your answer the same, Professor Turitsyn?

It is close. We need multidisciplinary, but, again, as far as I am concerned, if you ask the UK industry about problems, the answer will be different.

Professor Turitsyn, you said that we are behind the Netherlands. Is this an area that the UK can lead on?

Yes, because we have very strong, world-leading photonics centres. We have strong AI groups. We have a lot. We have semiconductor facilities and groups. We have all the components. We have this culture of research, which we need still to translate to the benefit of industry and Government.

For the time being, we can lead in decentralised AI using digital. This is what we could do—and we should.

In the interests of time, I will very quickly ask the same question that I put to the previous panel. The Government recently issued their AI hardware plan. How does that work for you?

I suppose you are talking about the AI hardware plan that was shared with me recently. There are some pockets of brilliance. I was happy to see so many aspects of it on, as I said, edge AI systems and on bringing companies in. There are some missing parts or gaps that I might bring up. For example, I have not seen a discussion about physical AI anywhere. We have a dire need to explore this area—for example, drones and collaborative AI, where devices talk to each other. The future of AI is physical, where it is coming out. For example, companies such as Nvidia are investing heavily in physical AI, and this is something I have not seen a discussion of. Also, with regard to bringing academics into this plan, I have not seen much in terms of using the innovations that we develop in labs to be able to support industry. This was also one of the gaps. A third one is cyber-physical security. There were some points about cyber-security in general, but cyber-physical security is as important. Again, I mentioned that the future of AI is physical, so it is going to go away from data centres and out towards the world.

It is a brilliant document. Surprisingly, for this Committee, low-energy consumption was not mentioned too much as one of the priorities. It was more about hardware and scaling. It is very important to add that we need multidisciplinary connections, which we discussed already. It is very important for new AI and for new generation. Also, the connection with commercialisation and industry could be stressed through some already-working instruments, such as knowledge transfer partnerships. They should be supported and scaled up, because that is where industry and, sometimes, small companies put their money, because they really believe in that. The university provides knowledge and transfers this knowledge to the industry. Also, I would like to mention the brilliant concept of CASE studentships, where 50% is coming from industry and 50% from Government, which is now focused on a list of very specific universities. Industry should, if it puts in its money, be allowed to choose academic partners, because there are good research groups beyond the Russell Group. If you want to help industry, ask where it wants to put in its money, rather than telling it.

I asked before about what skills were needed. Professor Kanjo, you talked about AI systems. If we have these studentships and are giving more support to industry, what undergraduate or master degrees or PhDs would support the future of neuromorphic photonics? What should people be studying?

It is an excellent question, because there are so many disciplines.

Just give me a few of them then.

It would be some combination of neuroscience, electronics, AI, computer science and photonics. I know it sounds very eclectic and strange, but maybe it is the future where some people and even businesses, I would say, can see how this can come together.

AI optimisation is very important for everything from data centres—

But there are no undergraduate courses in AI optimisation.

It is a big area. We spend ages trying to squeeze models into small devices.

I understand that, but what is the undergraduate degree or the master degree?

Tiny machine learning or edge AI.

We run several programmes. We applied for Erasmus Mundus to build neuromorphic computing as a centre of that. Of course, it is different and complex—

You are emphasising that it is really multidisciplinary.

Yes.

Bringing together silicon photonics and neuromorphic computing—you talked about the movement of data and embedding this hardware into wider systems—is there a use case for this for improving cyber-security or is that not relevant?

At the moment, because we are at a very low level of development for many of these things, it is very hard even to focus too much on cyber-security, but it is on our minds. You want these tools to work first before you think, “I now have something that might work. Can I think about the security of it?” We are trying to explore the feasibility of the technology at this stage, but security is on our minds. As I mentioned before, physical security in particular is a great concern that we have to look at.

With the new architecture for computing, it will take some time for cyber-attacks to catch up, one would hope. Thank you very much. We have gone somewhat over time, but that is testimony to how interested the Committee is in your work and also how fascinating your responses to our questions have been. This is a really exciting area. It is real science with real-world applications in the short, medium and long term. Thank you very much for joining us this morning.