Giovanni Acampora - Quantum Artificial Intelligence

excited to introduce Joavannia Kapankura Aampora full professor of computer science at the University of uh Naples and we're going to talk about quantum artificial uh intelligence and I will leave directly the word to him to introduce himself and the talk. Thank you very much.

Thank you.

Thank you for your kind introduction. First of all, I would like to thank the organizer for inviting me to this important event here in uh Naples and today it's my pleasure to introduce you with uh uh some research uh results about quantum artificial intelligence. So I would like to introduce you with uh with these topics in particular with the concept of artificial intelligence, the concept of quantum computing and I would like to to uh show how these two technology can be merged together in order to create this innovative research field named quantum AI. I quantum artificial intelligence. So what is artificial intelligence?

So today everyone discuss about AI. We use AI every day with chat GPT and with other uh mobile apps and so on. But from the computer science point of view, what is AI? AI is the capability of computers to emulate the abilities of biological entities such as for instance human beings. Okay.

in understanding, learning, making decision and adapting to their surrounding to the environment they live. And in particular, in order to understand this definition, we can make some examples. For instance, example, an example is related to autonomous driving. Okay, an autonomous car use computers and AI to see the road, learn from experience, make decision like when to stop or turn, and adjust to what's happening around it. just like a human driver would.

So this is an example of AI tool the autonomous car at the same time similarly we can introduce the concept of robot vacuum. Okay, a robot vacuum is an AI tool because it is able to detect obstacles, learn the layout of your home, to define the map of the environment, decide where to clean next and adapt to the change in its environment just like a person would when tidying up a room. Okay, so they are emulating people in performing some activities. And now maybe everyone knows chat GPT. Chat GPT is another example of artificial intelligence tool based on large language models that is able to understand the language, learn from patterns in text, decide how to respond and adapt to its answer based on the conversation.

Okay. So this example are um implementation of the definition I shown in the first in the first slide. Okay. So now we know the definition, we know the application but how is it possible to implement this application on our computers, on our mobile and so on. This is possible thanks to the concept of machine learning.

Okay, artificial intelligence can be viewed as a kind of group of different methodologies. We have reasoning methodologies, problem solving methodology, optimization methodologies. But currently the most used and the most important part of artificial intelligence is named machine learning. That is a methodology of artificial intelligence that allow computers to learn from example and improving their performance step by step without needing instruction for every task. Okay.

So thanks to machine learning a computer is able to learn from experience and this experience is represented by past data. Okay. And currently the the most used approach of machine learning are neural networks. Okay. And the neural networks are the main tool behind today's smart computer.

Behind chat GPT, behind every kind of AI system, there are smart there are neural networks. Okay. And what is the goal of a neural network? The goal of a neural network is to emulate the behavior of human brain. Okay.

Human brain is a collection of cells named neurons interconnected by synapses in a very simple uh and abstract way. And so during our experience during the human beings experience we are able to train our neurons to learn how to perform some activities. Okay. And so it's thanks to the experience that human beings are able to train their brain. And at the same way a computer can learn from experience thanks to the idea of neural network.

Okay. This is a very simple neural network emulate the concept of brain. The circles are the neurons and the arrows are the the synapses among the different neurons. And moving from the input to the output through the hidden layers, we are able to perform some calculation. Okay, our neural network that is a way to simulate the brain is just a mathematical computation.

Okay. As a mathematical computation that allow us to perform complex activity such as for instance recogni recognize an image such as in this case we have a cat. We would like to recognize this cat. We use the different pixel of the image as input of this neural network. we move from left to right by means of mathematical computation and at the end in the output layer we uh we get the probability that the image is related to a bird, cat or dog and as you can see here we have the 99% to be a cat.

Okay, but it is just a mathematical computation. And what is the one of the main ingredients of a neural network that allow a neural network to behave to behave in a correct way? It is the concept of weight. This u w variable in each layer. These are the weight of a neural network and the behavior of a neural network depending upon the value of the weight.

So it is very important to train a neural network to set up the value of the weights in a proper way in order to allow the network to have the right the right behavior. And this is the most important part of an AI tool. Okay. When we discuss about AI, the complexity of AI from the computational point of view is related to the training of the eye. Okay.

So when we train a network the result of the training the value of the weights can be organized in a matrix okay in a collection of number. So the neural network from a graphical point of view sorry sorry from a graphical point of view can be considered as a graph. Okay, a collection of nodes interconnected by arrow. But from the mathematical point of view, a neural network can be represented by using the weight matrix. Okay, that is the matrix that we need to learn in order to allow the neural network to work properly.

Now what is the problem that current AI system are very useful are very good but they are very complex. what means they are very complex that the number of weights that compose this AI uh network is very very high. Okay. In particular if we consider system like chat GPT the number of weights that ch GPT use to work properly is between 500 billion and one trillion. So we need to train this very huge matrix in order to allow chhatt GPT to work properly.

As you know now I show you how a matrix related to a neural network is represented. Okay. But how this matrix is uh stored in a memory of a computer. This matrix is stored as a sequence of bits. Okay, you know that in the memory of a computer we have a collection of a binary digit zero and one.

Okay, and because each number each weight of a neural network can be represented by using for instance 32 bits. I need a very huge number of bits to to store an artificial intelligence model. Okay, because I have a very huge number of weights. To represent each weight, I need several bits. So in the memory of a computer, I need several bits to store a neural network.

And so this is the problem that we would like to face. We need a way to train artificial intelligence system composed of billions on trillions of uh weights where each training step involve performing several operation on this number. We can use we can use a specialized hardware such as GPU and TPU to speed up this process. But however there there is a a big problem. The big problem is that current computers cannot be farther improved.

current computer are uh uh strongly related to the Gordon Moore and uh because the size of a transistor is becoming very very small is is it no more no more possible to further improve the performance of computers and so how we can deal with the next generation of AI system if current computers cannot be improved in the future. The solution is in quantum computing. Quantum computing is a very complex topic. And in order to explain you what is quantum computing, we will uh deal this uh topic by using maybe the the most used example that is the Schroinger cat. Okay.

And for this reason this is light. A cat can save us because thank to this phenomena related to the Schroinger cat we will be able to build computer more powerful than the classical computer we use today. Okay. What is quantum mechan? Quantum computing is different from classical computing because it is based on a different physics.

Current classical computers are based on classical physics. Quantum computing is based on quantum mechanics and quantum mechanics is that part of physics that describe how behave uh tiny particles such as electron uh proton and uh and so on and why quantum mechanics is different from classical physic because this particle can have some behavior very strange. Okay, for instance, this particle can be in different position at the same time. Okay, and this is not possible in the classical world where we live. I am here and I cannot I I cannot be simultaneously in another place.

But when we deal with quantum mechanics and with the very teeny particles so it's is it is possible to have this phenomena related named the superp position that allow these particles to leave simultaneously in different position. Okay. And starting from this uh discussion we can introduce the concept of shrodinger cat. Okay. If we move this quantum mechanical mechanical vision to the classical world, we can say that if we put a cat in a box in a particular in a particular experimental configuration.

So this cat can be simultaneously dead or alive. Okay, at the same time. Obvious obviously the experiment is not so simple. It depend upon some configuration and so on. But in order to explain here we have a cat in a box that is simultaneously dead and alive.

Okay. But if we have two boxes with two with two cats we have different configuration. We can have simultaneously that the two cats are dead and dead and alive and dead alive alive. So with two box the number of possible configuration is four. With three box is eight.

with four box is 16. So for each box added I double the number of possible configuration of the different cats in the different boxes. Okay. And moreover these cats can be connected among them. I can create a kind of connection between the cat in a box with a cat into another into another box.

And if I perform some operation on a box increasing the probability that the cat is dead at simultaneously I change the probability of the other cat in the other box. So a single operation can affect two cats simultaneously. So I am now introducing the concept of uh parallel computation. Okay. Because by means of superposition I have the different cats dead and alive simultaneously.

By means of the entanglement they can interact among them without communication. Okay. And so if we have n cats we have 2 to n possible configuration. And so we have an exponential quantity of information to store by means of only n boxes. Okay.

And this is one of the main benefit of quantum computation because the quantity of information that I am able to store in a quantum device is exponential with respect to the size of the device. Okay. So in a classical computer if the classical computer is uh composite of uh a single register storing n bits I can store only n bits in a quantum computer uh composite of a single quantum register composed of n cqits as we will see later I am able to store 2 to n cubits. So I am increasing the quantity of information that I can store and also the quantity of information that I can manipulate. Okay.

Okay. And so these are this is just a kind of recap of the the topic that I have discussed where we are replacing the concept of cats with the concept of cubits. So we are removing the cat from the box and we are introducing a bit in the box. The bit in a classical word can be zero or one. When I put the bit in the box, it can be zero and one simultaneously.

Okay? And I can relate different bits among them. When I act on a bit, the other bit is affected by this modification. Okay? And this bit live in this superposition of states till I open the box.

When I open the box, these cubits collapse in a classical register. Okay? In a classical configuration. So in general what is the impact of computation of this new configuration is that thanks to superposition a quantum computer with n cubits can explore 2 to n possibilities at once. Okay, instead of trying them one by one like classical computer and thanks to the entanglement, cubits can work together in a coordinated uh way implementing some uh computational operation that are not possible for classical for classical computation and this is the way uh why uh quantum computer can potentially solve certain problem much faster.

Okay, I would like I I like to say potentially because we have so far no theoretical proof that quantum computers are better. We have some example where quantum computers work better than classical computer but there is no a theoretical proof. Okay, only just example but maybe in the future uh computer science genius can implement a classical algorithm better than a quantum algorithm. And we have to close everything about quantum computation. But when in computer science when the time passes, yeah, we are confident that uh this hypothesis could be true.

And the idea of quantum computers, the goal of quantum computers is to achieve the quantum advantage. A quantum advantage means doing something with a quantum computer that a classical computer cannot do in a reasonable time in a reasonable time. Okay. So just to to complete this slide there is a very simple operation that a classical computer is not able to perform in a reasonable reasonable time that is the factorization. You know what is the factorization of a number?

If I have six the number six uh the factor of six are three and two. So I would like to have in a very simple way the list of number that multiplied among them return the original number. Okay we we learn at elementary school maybe at uh secondary school to how to perform this operation and it is strange but this operation is very hard for a computer. It means that you have a big number and you would like to know what are the number that multiplied among them return this big number a classical computer can uh use thousand years to to compute and indeed when you use WhatsApp for instance your messages are encrypted thanks to an algorithm based on this problem. So the encryption of our messages are hard to be decrypted because this problem is hard for a a classical computer.

And this is one of the main reason why quantum computer are very in interesting today because with the quantum computers maybe we we will be able to break this uh encryption algorithm on which our app are basing are based today. Okay. Indeed uh there exist an algorithm named Shor's algorithm introduced in 1994 that allow to perform the factorization in a polomial time. Polomial time means good. Okay.

And this is the strong interest in quantum computation is related to the its capability to solve this problem and this problem sometimes like this time are related to cryptography. Okay. And indeed in the United States they are there is the national institute for standard that is uh intro in introducing the postquantum cryptography algorithm. Uh yeah because this algorithm exist but there are no computer able to run them. But when in maybe 15 years 20 years we will have computers able to run this algorithes to perform the encryption of data.

Now what is the goal? Why I would like to merge together AI and quantum computing? Because we can use a kind of metaphor. I would like to replace the cat in the box with a net. I would like to put in the quantum box a neural network.

Okay. If I am able to put a neural network in the box, so I can use the superposition in order to have different configuration of the neural network simultaneously uh embodied in the in the box and I can try to speed up the training working on this parallel version of uh superposition version of the neural network. So this is the the idea that I can try to speed up the training of of neural network by using the quantum principles such as superp position and entanglement. And in this way also the next generation of AI tools composed of trillion of trillion of trillion of weights uh could be trained efficiently. So this is our our goal.

For instance, just to make an example, if we consider the current the current uh version of chat GPT, chat GPT is composed as said before uh by a number of weights uh between 500 billion and one uh trillion. And in order to store this number in a classical computer, we need tens of trillions of bits. Okay? But by using a quantum computer, each uh n cubit register can represent as I said before an exponential number of value. And for this reason, a computer a quantum computer equipped with only 40 cubits will be able to store tens of trillions of bits.

Okay, I need only 40 cubits to store all this information. Okay. Today we have at University of Naples we have a computer composed of 40 cubits. Uh IBM uh provide by using its cloud uh system computer composite of 127 cubits. But however even though we have this computer this computer as we will see later does not work properly.

Okay they have a good number of cubits but they are not good. I explain now why. Okay. So this is a quantum computer in particular. This is a quantum computer based on superconducting tech technologies and uh uh we have an example of this computer at department of physics of university of of Naples.

There are different technologies. Now another problem with the quantum computation. Now uh if you uh if you see TV uh news uh every week there is a news saying uh Napo is the first city with a quantum computer. Torino is the first city with a quantum computer. Salerno will be the first city with a quantum computer.

Every city will be the first with a quantum computer. In particular, these computers are based on superconducting cubits. That is currently the most uh uh suitable approach to build this kind of computer. But because the technology is not mature, there are also other way to try to design computers such as for instance ions or photonic quantum computing that are based on the concept of light, spin cubits and topological cubits. So there are different technologies we can try to use to uh implement a working quantum uh computer but so far no technologies is is useful to build a a working quantum computers.

What means that this computer uh are not able to work in a proper way. It means that we are leaving the so-called NISK era. Okay. We are in the NISK time. NISK means noisy intermediate scale quantum.

Okay, it seems hard to understand but it's very simple. It means that the computer are small and noisy. What means noisy? Noisy means that when you perform a computation, the result is wrong. Okay.

If you ask is a in a very simple way, you don't ask to a quantum computer the sum between three and two. If you ask the sum between three and two to a quantum computer, sometimes it return four, sometimes six, sometimes three, sometimes five and so on. So there is this noise in computation that not allow us to perform the computation in a right way. Moreover, the size of computer is small. Okay, the number of cubits that currently equip uh quantum computers does not allow us to face important problem.

But we are working on we can say toy problem introduction problem and so on. In order for instance before I discussed about the short algorithm the algorithm to factor to factoriize a number. Okay. In order to implement that algorithm in a correct way we need 1 million logic cubits. 1 million.

Now we have 40 50 100. So we have to arrive to one one million logic cubits. not noisy only when we will be we will be able to build this kind of computer we will be able to break the encryption algorithm of WhatsApp and uh and so on. So we are far indeed two three weeks ago the CEO on video is a top company in in uh in computation say that in his vision we will have a practical quantum computers in 15 30 years. Okay, we will be able to build a computer able to solve a practical problem in 15 30 years.

Maybe this is not this is a pessimistic vision but maybe 10 years is uh a good time. Okay. And then as I said currently with this computer with the N computer we are able to run small to medium scale quantum algorith. So we are able to solve a little problems. Okay.

Currently classical computer are better than quantum computers in solving the problem. And we can explore how for instance quantum computing can be used to improve artificial intelligence. That is our goal in our laboratory in our department. And another important topic I think my vision is that quantum computer cannot work alone. Okay.

I see a quantum computer as a kind of side car. You have to put a quantum computer together with a big a powerful classical computer to solve a particular problem. So in my vision uh will not use quantum computers like we use our phone today or our laptop today but that machine will be machine to solve very hard problem in particular application domain and so on. So you have to imagine these computers in high performance computing environment uh working together with the classical computers and now uh there is another challenge. Okay, I say that quantum computing could be a benefit for AI because with quantum computers we can try to store big artificial intelligence model in efficient way.

Okay, but however there is also the other direction. The question is classical artificial intelligence can be used to improve the performance of quantum computers because qu uh current quantum computers are not so good. Okay, but maybe using AI we are able we will be able to to solve some issues related to this computer and in particular for instance we could try to reduce the noise in computation thanks to AI or we can try to perform the so-called compilation of quantum programs in a more efficient way thanks to AI. So there is when we discuss about quantum AI we have to consider a birectional uh interaction because quantum computing can support AI but currently AI can support quantum computing in uh in working better. Okay.

Okay. Now after this uh vision about integration of quantum computing and AI, I would like to show you what we done in our laboratory in both direction. We worked uh on the direction quantum for AI and on the direction AI for quantum and uh in particular okay this is just the name of our laboratory is named quazer quantum computing and spar system funded by me in 2017 at university of nap in of nap federico secundo and this is maybe one of the first maybe the first laboratory in Italy investigating about uh the integration of quantum I I yes know it's the first like yeah maybe maybe tomorrow there will be another first laboratory in Salerno maybe this is I think this is the first just for a casuality because when I uh come back in Italy from UK I moved to the physics department and I had I was lucky to as a computer scientist to move to department of physics and to explore this new topic uh uh before than than other okay is one of the the first and we would like to design intelligent system that combine the power of quantum mechanics with the reasoning capabilities of AI. So this is the direction quantum for AI and at the same time we would like also to work on AI for for quantum For instance, in 2021, we introduced the first genetic algorithm able to be run on a quantum computer. A genetic algorithm is an optimization algorithm inspired by nature.

So we solve optimization problem at the same way how uh population evolves following the Darvinian principles. Okay. But this algorithm were implemented on classical computers before in 2021 we introduced the first quantum version and it work in a good way. We were awarded by IBM for this uh project and thanks to this pro award we were able to to use the bigger computer of IBM to test the the algorithm. Then we have introduced the first quantum fajy system.

What is a fajy system? I will try to explain in a very simple very simple way is a way to program computers classical in particular by using linguistic rules. Okay, I can pro yeah like like that I have a rule if the temperature is ike and the humidity is moderate then set the fan speed to medium ike. Okay. Thanks to fatsy logic and fuzzy system, I can program a computer by using the English words.

Okay. And we say it's possible to make the same on a quantum computer. And there is a benefit. Yes, we prove it that it's possible and there is a benefit. What is the benefit?

The benefit is if you consider this rule. If we increase the number of variable in this rule for instance temperature, humidity, I add more variable. The number of rules that I can compose grows exponentially. And a classical computer need an exponential time to run this rule. But on a quantum computer, we need just one execution.

And in a polomial time, we are able to perform this operation. And so we proved that is it possible to program a quantum computer by using linguistic rule and you have a computational benefit thanks to this system. This system is not only a theoretical system. We move to the CERN in Genev and uh we control it up particle accelerators by using linguistic rule and a quantum computer. Okay.

And uh this is the the picture from the control center of the particle accelerator where our PhD student past PhD student now is a young researcher uh is controlling a particle accelerator thanks to our algorithm and the same algorithm has been applied to the control of a smart city. So we can control the different part of a smart city in a linguistic way by a quantum computer. We participated to a uh the Fujitsu quantum award and we won this uh award thanks to this project. So uh Fujitsu awarded us with uh yeah $10,000 for our research. We were awarded in in the Netherlands at the Fujitsu European Fujitsu Center.

Yeah. Because it's one of the the practical application of quantum computation. Obviously as said before this is just a theoretical result because current quantum computer are not able to run properly our algorithm but in an ideal case if we suppose that the computer is ready it work good. Okay. And Fujitsu awarded us.

Then we also we are also developing software to allow people to train AI models on quantum computers. And the software is named Evovak is named Evovak is a evolutionary algorithms for variational quantum circuits. Okay. where we use genetic algorith classical genetic algorithm to to uh model quantum artificial intelligence system and it works in a in a good way. It is a it is strongly used and starting from that research we implemented an optimization algorithm for quantum computers awarded as the best journal paper just one week uh two weeks ago in uh in row.

Then now the previous slide are related to quantum computing for AI but we also worked on AI for quantum computing. We use a clustering algorithm to collect the different kind of noise that a quantum computer is uh uh that a quant uh that a quantum computer software. Okay. And uh by means of this algorithm based on unsupervised machine learning, we were able to improve the performance of IBM computers. Okay.

Better than the algorith the IBM algorithm to make the same activity. Okay. And this this uh research was another best paper award uh at a fat 2021 in Luxmbborg and we are also working on the as I said before on the compilation. Yeah, when you run a a quantum algorithm, you have to compile this algorithm in a form that is compliant with the processor where you would like to run the algorithm. And we use artificial intelligence to find the right transformation to run the algorithm in a good way.

Then yeah um beside the the research activity uh we also we are also working on editorial and conference activity in particular in 2029 2019 sorry 2019 we introduced the first the first journal in quantum artificial intelligence and currently is a kind of uh is a kind of the main forum room for scientists working with this uh with this uh topic and in uh November we are organizing a conference here in NALE is the first conference sponsored by ILE E that is the top uh entity top the top body for engineering in the world on quantum artificial intelligence from November 2 to five. Okay. And uh we are able to organize this conference because in 2022 we were awarded by the e computer society with the emerging technology award. So they provided us with uh some fund to to create a research network in quantum artificial intelligence and starting from this research network now we are able to organize the first conference on quantum AI in the world. The next edition will be in Nottingham and in Athens but the first is in Napoli this year.

Okay. Okay. So in conclusion uh why quantum computer are useful now because classical computer could be not able to further improve uh to support uh the next generation of AI tools. Okay. because as said before uh we will be not able to make transistors smaller and smaller.

Okay. And for this reason uh it is important to think an alternative computational paradigm such as quantum uh quantum computing because thanks to its feature such as superposition and entanglement it could provide a computational advantage in AI. But as said before I think that we need 10 around 10 years in order to have a kind of practical quantum AI system. But we are working to try to speed up okay the the process. Okay.

So thank you. Thank you very much.