Repo for the code used to generate the data in the paper I co-authored (first author) Sample efficient graph classification using binary Gaussian boson sampling forthcoming in the journal Physical Review A. Avaialbe here: https://arxiv.org/abs/2301.01232v2
A boson sampler is a near term quantum computer that is strongly believed to be classically hard to simulate (see Scott Aaronson & Alex Arkhipov 2011).
The basic idea of the quantum machine learning algorithm is to encode the adjacancey matrix of a graph into a Gaussian boson sampler that measures photon detection events using thereshold (binary) detectors which click when they detect any photons as opposed to photon number resolving detectors which count exactly how many photons are detected.
After the matrix is encoded into the device we sample the device S times. Using those samples we construct a feature vector that contains useful information about the graph encoded.
In our case the features are related to the number of perfect mathings of all possible subgraphs of the encoded graph.
The probability of a detection event, denoted
where
Folders are named after the model used to benchmark the new feature vectors. NN being the multi-layer perceptron model, GBS being the Gaussian boson sampling model, RW being the random walk kernel, SP being the shortest path kernel, SM being the subgraph matching kernel and GS being the graphlet sampling kernel. GBS-mu contains the results for the feature vectors whose featrues correspond to the probabilites of detection events with the same number of detector clicks. GBS-nu contains the results for the feature vectors whose featrues correspond to the probabilites of the 32 possible detection outcomes of a 5 mode binary GBS device. The folder nu-32 also contains the results of principal component analysis of the featuers extracted by one of the coarse-graining strategies presented in the paper.