Quantum Deep Equilibrium Models

Authors: Philipp Schleich, Marta Skreta, Lasse Kristensen, Rodrigo Vargas-Hernandez, Alan Aspuru-Guzik

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply QDEQs to find the parameters of a quantum circuit in two settings: the first involves classifying MNIST-4 digits with 4 qubits; the second extends it to 10 classes of MNIST, Fashion MNIST and CIFAR. Our code is available at https://github.com/martaskrt/qdeq. Section 3 Experiments, Section 4 Results.
Researcher Affiliation Academia Philipp Schleich Department of Computer Science University of Toronto Vector Institute, Marta Skreta Department of Computer Science University of Toronto Vector Institute, Lasse B. Kristensen Department of Computer Science University of Copenhagen, Rodrigo A. Vargas-Hernández Department of Chemistry & Chemical Biology Mc Master University, ON, Alán Aspuru-Guzik Department of Computer Science Department of Chemistry University of Toronto Vector Institute
Pseudocode No The paper contains figures illustrating circuit diagrams (Fig. 2, Fig. 3) and a process diagram (Fig. 1) but no explicitly labeled 'Pseudocode' or 'Algorithm' blocks with structured steps.
Open Source Code Yes Our code is available at https://github.com/martaskrt/qdeq.
Open Datasets Yes First, we consider MNIST-4, which consists of 4 classes of MNIST digits (0, 3, 6, 9) (Deng, 2012). Fashion MNIST (Fashion MNIST-10) (Xiao et al., 2017). Finally, we tested our setup on natural images with CIFAR-10 Krizhevsky et al. (2009).
Dataset Splits Yes For all datasets, we used default train/test splits2 and randomly split the training set into 80% train, 20% validation.
Hardware Specification Yes Runtime was calculated over 100 epochs on a NVIDIA RTX 2070 GPU.
Software Dependencies No As mentioned, all results were generated using the torchquantum framework (Wang et al., 2022a). This mentions a framework but lacks specific version numbers for it or any other software dependencies (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes We train the implicit models using a Broyden solver for at most 10 steps. For optimization, we use Adam (Kingma and Ba, 2014) and cross-entropy loss. We trained each model for 100 total epochs (i.e. if we first pre-trained using x warm-up epochs, we then trained using the implicit framework for 100 x epochs) (for CIFAR-10, we only trained for 25 total epochs since we found it to converge faster). We selected hyperparameters using the validation set; see Appendix E. Appendix E (Table 5) provides specific hyperparameter values.