Gradient Estimators for Implicit Models

Authors: Yingzhen Li, Richard E. Turner

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The efficacy of the proposed estimator is empirically demonstrated by examples that include gradient-free MCMC, meta-learning for approximate inference and entropy regularised GANs that provide improved sample diversity.
Researcher Affiliation Academia Yingzhen Li & Richard E. Turner University of Cambridge Cambridge, CB2 1PZ, UK {yl494,ret26}@cam.ac.uk
Pseudocode No The paper describes methods through text and mathematical equations but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes Implementation is released at https://github.com/Yingzhen Li/Stein Grad.
Open Datasets Yes We take from the UCI repository (Lichman, 2013) six binary classification datasets (australian, breast, crabs, ionosphere, pima, sonar)
Dataset Splits Yes We use ionosphere as the validation set to tune ζ. The remaining 4 datasets are further split into 40% training subset for simulating samples from the approximate sampler, and 60% test subsets for evaluating the sampler s performance.
Hardware Specification Yes All the methods are timed on a machine with an NVIDIA Ge Force GTX TITAN X GPU.
Software Dependencies No The paper mentions using 'Adam optimiser' but does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions or specific library versions).
Experiment Setup Yes For the training task... The step-size ζ is selected as 1e-5... The learning rate is set to 0.001, the number of epochs is set to 500... The minibatch size is set to K = 100. Learning rate is initialised at 0.0002 and decayed by 0.9 every 10 epochs... The selected γ and α values are: for KDE estimator approach γ = 0.3, αγ = 0.05, for score matching estimator approach γ = 0.3, αγ = 0.1, and for Stein approach γ = 0.5 and αγ = 0.3.