First Steps Towards Quantum Artificial Intelligence (Quantum)

New research by a team from Los Alamos National Laboratory shows that Quantum Convolutional Neural Networks (Qcnn) do not have the Barren Plateau problem. According to Nicolò Parmiggiani (Inaf), the two theorems demonstrated in the study published in Physical Review X are very important for guaranteeing the applicability of Qcnn and for the future of quantum neural networks.

The convolutional neural networks running on quantum computers have generated considerable hype for their potential to analyze quantum information can do better than classical computers. However, the application of these neural networks with large data sets has always been a problem, because of a kind of “Achilles heel” known as Barren Plateau , or barren plateau . But now new research led by a team at Los Alamos National Laboratory , published in Physical Review X , appears to have managed to overcome this.

Media Inaf interviewed Nicolò Parmiggiani , researcher at the National Institute of Astrophysics, machine learning expert and winner of the first edition of the National Award for research on big data and artificial intelligence  for his studies on machine learning technologies applied to analysis of the data produced by the Agile space telescope .

Parmiggiani, can you explain to us what this new study is about and why it gives us hope for the analysis of complex systems?

«The article describes a quantum neural network (Qnn) model based on the Convolutional Neural Network (Cnn) architecture used in classical machine learning . CNNs are a particular architecture of neural networks in which (artificial) neurons are connected to each other with a pattern inspired by the structure of the animal visual cortex. These models have obtained successful results in various different fields, including astrophysics where they are often used to analyze two-dimensional images and solve classification or object detection problems . We, for example, as an Agile team have developed a CNN model to identify Gamma-Ray Bursts within the counting maps produced with the Grid instrument on board the satellite ».

So not only classical, but also quantum machine learning. How are the two different?

“In recent years, a lot of research has been done to combine the advantages of quantum computers with neural networks. Quantum neural networks are models of neural networks that rely on quantum mechanics to develop more efficient algorithms by exploiting the properties of quantum information. The article  describes the architecture of the Quantum Convolutional Neural Networks (Qcnn) which use the same principles of the classic CNN architecture to analyze a quantum state instead of classical inputs, such as images. The main purpose of the Qcnn is to be able to analyze very complex quantum systems that would not be tractable with machine learningclassic. For example, a problem that can be analyzed with the Qnn is the many-body problem ».

Nicolò Parmiggiani, researcher at the INAF OAS in Bologna © Inaf

Are there any limits to this approach?  

«Despite the enormous potential, quantum machine learning has some problems, including the so-called Barren Plateau , similar to the Vanishing Gradient problem that occurs in classic machine learning . During the training phase of very deep classical neural networks (with many layers ), the optimization gradient of the parameters that make up the model can become very small, not allowing the updating of the parameters of the neural network that is not trained correctly. A very similar situation can also be found in the NQF ».

Why is this problem visually traced back to a plateau , i.e. a plateau?

“The plateau metaphor, used in quantum machine learning , arises from the fact that if the algorithm is to minimize a cost function (to optimize the parameters) it will try to identify the direction (the gradient) to reach the valleys whose function cost is lower. However, it may happen that the algorithm is unable to find a direction to go and therefore finds itself in the so-called plateau where all directions seem the same or with very small differences.. In this situation, the NQF often cannot be trained efficiently and is therefore unusable. This problem can get worse as the number of parameters increases and therefore does not allow to exploit the advantages of quantum computers to analyze large complex systems ».

How important is the study published by the Los Alamos National Laboratory team ?

“The published study shows that the Qnn, which are Qnn built with a particular architecture, do not have this problem of the Barren Plateau but can be managed with a complexity that grows in a polynomial way with respect to the size of the system and which guarantees the possibility of training the model . In fact, one of the main advantages of using the Qnn is that of being able to analyze with polynomial complexity systems that would have an exponential complexity in classical machine learning and therefore obtain a clear computational advantage. The two theorems demonstrated in this article are therefore very important to guarantee the applicability of the Qcnn and for the future of quantum neural networks ».

Featured image: New evidence that it is possible to secure the training of certain quantum convolutional networks paves the way for quantum artificial intelligence to aid in the discovery of materials, as well as in many other applications. Credits: Lanl

To know more:

Provided by INAF

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s