Aprendizaje automático basado en computación cuántica. Clasificación de imágenes mediante Kernels cuánticos

Abstract

Recientemente se ha alcanzado la era de los dispositivos NISQ (Noisy Intermediate-Scalce Quantum), dispositivos que, pese a seguir siendo susceptibles al ruido, pueden mejorar sobradamente a los computadores clásicos en ciertas tareas. En este proyecto se trata de obtener una simulación de un modelo que mejora de forma teórica las capacidades de las soluciones clásicas existentes. Una idea recurrente dentro de esta disciplina es hacer uso de circuitos cuánticos como redes neuronales, debido a la similitud en su estructura. De esta forma se pueden optimizar los parámetros con métodos similares. En este proyecto se ha resuelto un problema de clasificación binaria supervisada de imágenes. Para ello se ha utilizado el método del núcleo o <<kernel>> cuántico, que es similar a los kernels clásicos pero siendo ejecutado en un dispositivo cuántico ya sea real o simulado. La computación cuántica aporta espacios de Hilbert de alta dimensionalidad donde separar las distintas observaciones del problema. La mejora de la eficiencia teórica con respecto a los métodos clásicos ya ha sido demostrada.

Recently, the NISQ (Noisy Intermediate-Scalce Quantum) era devices has arrived, although devices are still susceptible to noise, they can improve on classical computers in certain tasks. The aim of this project is to obtain a simulation of a model that would improve the capabilities of existing classical solutions. A common idea within this discipline is to use quantum circuits as neural networks, because of the similarity in their structure. In this way, parameters can be optimised using similar methods. In this project, a supervised binary image classification problem has been solved. For this purpose, the quantum kernel method has been used, which is similar to classical kernels, but being executed on a quantum device either real or simulated. Quantum computing provides high-dimensional Hilbert spaces in which is possible to separate the different observations of the problem. The improvement in theoretical efficiency over classical methods has already been demonstrated. The classification problem solved in this case is the traditional separation of images of dogs and cats. It is a hybrid approach where a classical convolutional neural network extracts some features from the different images that are evaluated by a quantum circuit that raises the dimensionality of the data performing as the kernel function and then, the images will be classified with support vector machines. The growth of quantum technologies has encouraged some companies to develop free libraries with which to develop projects and simulations on real quantum computers, as in the case of Pennylane developed by Xanadu.