What quantum computing will change in AI

The exponential increase in computer power that takes shape opens up new perspectives in terms of machine learning.But without revolutionizing the concept.

With the advent of the quantum computer, IT will enter a new era.This future generation of supercomputer will multiply the computing power by two at each new qubit (or quantum bit).The promise ?To be able to process treatments in a few seconds requiring traditional supercomputers several thousand years.But to achieve this result, it will first be necessary to create sufficiently stable qubits.Which is not yet won.Amazon, Google, IBM, Microsoft, as well as Oat in France are in the race.In parallel, all already anticipate the next step: the development of optimized algorithms for the new architecture.And among these applications, artificial intelligence is in good place.

""Thanks to the power provided by quantum computer, it will become possible to cause models of machine learning on gigantic learning bases"", underlines Florian Carrière, senior manager in charge of emerging technologies at Wavestone.From there, a quantum algorithm based on a probabilistic method could identify connections formerly impossible to discern.""An automatic language processing model like GPT 3 already has 175 billion learning parameters.We will be able to go even further. Ce qui va permettre de passer un nouveau gap en matière de traduction complexes ou d'analyse de sentiments"", avance Cyril Allouche, en charge du programme de R&D quantique chez Atos.

Billions of billions of parameters

Ditto for computer vision.In this area, convolutive neural networks will be able to make up billions of billions of parameters, notably benefiting from the rebuilt networks.The recognition of forms and scenes, in the autonomous vehicle for example, will reach a degree of finesse and unprecedented precision.Alongside the size of quantum memory, the advantage will obviously lie in the acceleration of the learning speed.

Beyond the use of already existing artificial intelligence models, could quantum infrastructure give birth to new kind learning structures?""Use quantum entanglement to create a new type of network of neurons, in which weights would not be real values but values in superposition, does not present more.Researchers and manufacturers have reached a consensus on this point, ""replied Cyril Allouche.

As for existing machine learning models, they will be easily applicable to a stable quantum environment with a minimum number of qubits.""The majority of research published on the subject are intended to convert current learning algorithms in order to run them on quantum environments,"" insists Xavier Vasques, CTO and Distinguished Data Scientist at IBM France.""Generative or gan antagonistic networks or Vector Machine support (SVM), are good example.""

In the case of SVMs, core functions (exponential, Gaussian, hyperbolic, angular, linear) solve problems of classification, regression or detection of anomaly within 2D or 3D space. ""Plus le volume de caractéristiques sera grand, plus le calcul sera coûteux en termes de puissance machine. On gagne donc à transformer ces fonctions en algorithmes quantiques pour accélérer le calcul"", explique Xavier Vasques.Another advantage of quantum computers: it makes it possible to define the hyperparameters of a model in a three -dimensional space more quickly by paralleling 3D calculations.

""Les premiers résultats que nous avons obtenus sur les algorithmes quantiques montrent qu'ils parviennent à détecter des paterns dans des données bruitées là où les algorithmes classiques n'en identifient pas"", constate Xavier Vasques. ""Pour observer la production du Boson de Higgs qui est extrêmement ténue, le CERN utilise des SVM quantiques pour détecter des micro-évènements révélateurs dans les mégadonnées produits par son accélérateur de particules. Ces machines à vecteur de support génèrent une classification du signal, du bruit de fonds…"" Des SVM quantiques qui parviennent au même résultat que les classificateurs développés par le CERN openlab en s'appuyant sur des méthodes classiques. ""Ce qui laisse présager de belles avancées au fur et à mesure de la progression de la recherche en matière de hardware quantique"", conclut Xavier Vasques.

The Nisq

From SVM to networks of convolutive neurons, quantum computers suggests an acceleration of research in multiple fields: genetic analysis, prediction of protein structure and discovery of new treatments, logistics optimization, product recommendation engine, detection offraud, chemistry, development of new materials...In industry, Airbus and EDF are engaged in quantum computer science research.Another very invested sector: finance.In this area, Barclays, Goldman Sachs and JPMorgan are on the ranks.Their objective is in particular to take advantage of the machine learning statistical learning to lead to stronger prediction models or refine the scoring of assets or credits.

While waiting to benefit from stable quantum machines, the first NISQ systems (for Noisy Intermediate-Scale Quantum) should be released by 2023.These are quantum machines counting between 50 and a few hundred qubits, but with too short stability to carry out certain operations. ""Les SVM par exemple ne fonctionneront pas en mode NISQ.Reinforcement Learning is more promising. Quant aux réseaux de neurones, ils nécessiteront des centaines de milliers voire des millions de qubits pour fonctionner dans ce mode, le coût d'encodage étant quadratique par rapport au nombre de paramètres à optimiser"", explique Cyril Allouche chez Atos.Quantum Deep Learning could nevertheless prove more robust to the noise of NISQ systems in image recognition.As often in computer science, the first quantum AI should not be without constraints.