Monotone operator equilibrium networks

Authors: Ezra Winston, J. Zico Kolter

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To test the expressive power and training stability of mon DEQs, we evaluate the mon DEQ instantiations described in Section 4 on several image classification benchmarks. We take as a point of comparison the Neural ODE (NODE) [8] and Augmented Neural ODE (ANODE) [10] models, the only other implicit-depth models which guarantee the existence and uniqueness of a solution. We also assess the stability of training standard DEQs of the same form as our mon DEQs.
Researcher Affiliation Collaboration Ezra Winston School of Computer Science Carnegie Mellon University Pittsburgh, United States ewinston@cs.cmu.edu J. Zico Kolter School of Computer Science Carnegie Mellon University & Bosch Center for AI Pittsburgh, United States zkolter@cs.cmu.edu
Pseudocode Yes Algorithm 1 Forward-backward equilibrium solving; Algorithm 2 Peaceman-Rachford equilibrium solving; Algorithm 3 Forward-backward equilibrium backpropagation; Algorithm 4 Peaceman-Rachford equilibrium backpropagation
Open Source Code Yes Code is available at http://github.com/locuslab/monotone_op_net.
Open Datasets Yes We train small mon DEQs on CIFAR-10 [17], SVHN [22], and MNIST [18], with a similar number of parameters as the ODE-based models reported in [8] and [10].
Dataset Splits No The paper mentions training and testing, but does not provide explicit details about validation splits, percentages, or sample counts needed to reproduce the data partitioning.
Hardware Specification Yes All experiments are run on a single RTX 2080 Ti GPU.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies (e.g., programming languages, libraries, or frameworks like PyTorch, TensorFlow).
Experiment Setup Yes For further training details and model architectures see Appendix E.