Neural networks are an Artificial Intelligence technique that mimics the behavior of the human brain. In neural networks, the neurons (or perceptrons) are connected to each other, forming a network (hence the name). Thanks to these connections, the information received is transmitted from one neuron to another, thus being able to process the data received. However, as they always say, “You can’t learn everything on the first try.” This means that it’s necessary to practice – do some trial and error – until you manage to improve the skill you want to learn. Let’s look at an example: wanting to learn archery. Your goal will be to hit the center of the target. The first time you try, the arrow will, most likely, hit a point far from the bullseye (or you might not even manage to hit the target at all). The second time, you’ll already have become familiar with how to hold the bow, and will be able to shoot the arrow more accurately. Next, you’ll find that by positioning your arm a certain way, you can better pin down the direction that the arrow will shoot. And you’ll try as many times as necessary until you achieve your goal. Neural networks are based on this learning process.

When a person learns a new skill, they have to practice it over and over again, making appropriate changes each time to hone their abilities. We can understand a neural network model with each neuron having a possible change to make. If we use the example of archery, we can say that one neuron is responsible for adjusting the shooting angle, another for the position of the arm, another for when to release the fingers... by adjusting each possible change for each neuron, we’ll finally be able to achieve the proper action allowing us to hit the target. This is why we say that the model is being “trained.” This is done with mathematical techniques that are beyond the scope of this article, but it’s important to remember the general idea: the neural network model can make certain changes until we get the desired result. With humans, once someone has learned how to shoot an arrow well, when they try the same thing again, they’ll do better than the first time, because they’ve already trained. The same thing happens with neural network models: once the model is trained, when given new data, it will know what to do with that data.

In what areas can we use neural networks? There are different types, each suitable for certain tasks. The areas of application vary from image detection using cameras to assessing value predictions based on prior data. In the case of Global Data Quantum, we use neural networks to predict the filling of dumpsters. We have data on the collection of dumpsters in Biscay. Using this data and neural networks, it’s possible to predict when a dumpster is going to be full or not. Therefore, during the waste-collection process, if we know that there is a dumpster that is going to be empty, the truck’s route will avoid them, thus saving time and fuel during the collection route.

Thus far we’ve explained what neural networks and their applications consist of, more or less, but where does Quantum Technology come into all this? Quantum Computing is a field that has been talked about a lot in recent decades. There’s great hype about this new technology, but the reality is that quantum computers won’t replace the conventional computers that we all know and love (which we’ll refer to as classical computers). Quantum computing will complement classical computing, thus being able to improve its performance on problems that are too complicated for today’s classical computers. This is where the technique known as Quantum Machine Learning comes in, which is an extension of Machine Learning that uses Quantum Computing. Thanks to this, we can now introduce the concept of quantum neural networks, which is one of the projects that I, as a Quantum Data scientist, am involved in at Global Data Quantum. As previously mentioned, we use neural networks to predict how full dumpsters are in order to optimize collection routes, but now let’s take a look at what happens when we combine classical techniques with quantum computing.

What’s the difference between a classical neural network and a quantum neural network? It must be clarified that, today, quantum neural networks are not purely quantum. This means that we use hybrid algorithms; that is, a combination of classical computing techniques with quantum computing ones. As we mentioned before, when we learn a new skill, we make certain changes until we achieve the desired goal. Quantum neural networks also make changes until they achieve their goals. The difference is that now we have to program our neural network so that it can be implemented on a quantum computer. This is achieved using quantum circuits, which are an approximation to Quantum Computing. Quantum Computing consists of doing computations (as its name suggests) while taking advantage of phenomena from quantum physics, such as superposition and entanglement. Superposition means that an object can be in more than one state at the same time (or take more than one value at the same time). Entanglement means that, even if we separate a system from one or more objects, what happens to one will happen to the other, despite being spatially separated. These types of phenomena can be reproduced by different techniques, but IBM’s quantum computers, on which we can run our programs, use quantum circuits, so I’m going to explain how we implement our quantum neural networks at Global Data Quantum.

Quantum neural networks, therefore, consist of a quantum circuit; specifically, a variational quantum circuit. The word “variational” tells us that the circuit has elements that can vary in value, which we will call “parameters.” The quantum circuit has elements whose value can vary, in the same way that we make changes to improve our archery technique. Being a quantum circuit, we’ll be able to take advantage of superposition and entanglement to process information faster when quantum computers become more sophisticated. The process followed by a quantum neural network is, therefore, as follows: first, the data is entered into the system and initially prepared to be transferred to the quantum circuit. Next, the data goes into the quantum neural network, which, as we said before, is a quantum circuit. This quantum circuit processes the information and gives us a result. This part of the process could be compared to a first attempt when practicing archery. Imagine that you’ve shot the arrow and it hasn’t even hit the target. Based on the result (you haven’t hit the target with your arrow) you’ll change, for example, the angle at which you shoot. Our quantum neural network will do the same thing. Based on the result, it will adjust some of the circuit’s values. This process, known as parameter optimization, is done in a classical way; that is, with a conventional computer. Once you’ve made your first attempt, the next thing is to repeat it, to see if it turns out better this time. The quantum neural network will do the same thing. It will repeat the process with the changed values. It gets a result, and then optimizes it again. And it will repeat the process until the desired results are achieved. Just like us learning archery, the quantum neural network also “learns.” The difference is that it won’t take months for the quantum neural network to learn its new technique. This is how we′re transferring the classical neural network technique to predict the filling of dumpsters via quantum computing.

Currently, the quantum advantage over classical computers has not yet been achieved, and further research into quantum computing techniques is needed. The applications of neural networks are very diverse, and this is an area in constant development. We could say that now our goal is not to learn to shoot an arrow, but to learn to implement neural networks.