QuAnt: Quantum Annealing with Learnt Couplings
Authors: Marcel Seelbach Benkner, Maximilian Krahn, Edith Tretschk, Zorah Lähner, Michael Moeller, Vladislav Golyanik
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We run our experiments on D-Wave Advantage5.1 (Dattani et al., 2019), an experimental realisation of AQC with remote access. We next experimentally evaluate Qu Ant. The quantitative results for graph matching, point set registration, and rotation estimation are reported in Tables 1, 2, and 3, respectively. |
| Researcher Affiliation | Academia | Marcel Seelbach Benkner Universit at Siegen Maximilian Krahn MPI for Informatics, SIC Aalto University Edith Tretschk MPI for Informatics, SIC Zorah L ahner & Michael Moeller Universit at Siegen Vladislav Golyanik MPI for Informatics, SIC |
| Pseudocode | No | The paper does not contain explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code and the new dataset are available at https://4dqv.mpi-inf.mpg.de/QuAnt/. Our code, which we will release, is implemented in Pytorch (Paszke et al., 2019). |
| Open Datasets | Yes | The code and the new dataset are available at https://4dqv.mpi-inf.mpg.de/QuAnt/. Data. We evaluate graph matching on the Willow object dataset (Cho et al., 2013), which contains labelled key points. We evaluate 2D point set registration on the 2D Shape Structure dataset (Carlier et al., 2016). For 3D rotation estimation, we evaluate on Model Net10 (Wu et al., 2015). |
| Dataset Splits | Yes | We use 5640 images for training, and test on 846 images. We use 500 shapes from various classes for training, and test on 50 shapes. We use 300 shapes from various classes for training, and test on 30 shapes from various classes. |
| Hardware Specification | Yes | We run our experiments on D-Wave Advantage5.1 (Dattani et al., 2019), an experimental realisation of AQC with remote access. |
| Software Dependencies | Yes | Our code, which we will release, is implemented in Pytorch (Paszke et al., 2019). We access the QA via Leap 2 (D-Wave Systems, 2022) using the Ocean SDK (D-Wave Systems, Inc., 2022c). |
| Experiment Setup | Yes | We use a multilayer perceptron (MLP) with L layers and H hidden dimensions, Re LU activations (except for the last layer, which uses sin (Sitzmann et al., 2020)), and concatenating skip connections from the input into odd-numbered layers (except for the first and last layers). L = Lgap + λunique Lunique + λmlp Lmlp, (5) where we set λmlp = 10 4 and λunique = 10 3 regardless of problem type. We use Adam (Kingma & Ba, 2014) with a learning rate of 10 3 for training. For graph matching on Rand Graph with k=4, we use a batch size of 141 and train for 150 epochs. For the point cloud experiments, we use a batch size of 32 and train for 20 epochs. For Qu Ant, Diag and Pure variants, we experiment with all combinations of the numbers of layers L {3, 5} and hidden dimensions H {32, 78}. When solving a QUBO on a QA, we anneal 100 times and pick the lowest-energy solution. |