Discretization Drift in Two-Player Games

Authors: Mihaela C Rosca, Yan Wu, Benoit Dherin, David Barrett

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To understand the effect of DD on more complex adversarial games, we analyze GANs trained for image generation on the CIFAR10 dataset. We follow the model architectures from Spectral-Normalized GANs (Miyato et al., 2018). Both players have millions of parameters. We employ the original GAN formulation, where the discriminator is a binary classifier trained to classify data from the dataset as real and model samples as fake; the generator tries to generate samples that the discriminator considers as real. This can be formulated as a zero-sum game: E = Ep (x) log Dφ(x) + Epθ(z) log(1 Dφ(Gθ(z))
Researcher Affiliation Collaboration Mihaela Rosca 1 2 Yan Wu 1 Benoit Dherin 3 David G.T. Barrett 1 1Deep Mind, London, UK 2Center for Artificial Intelligence, University College London 3Google, Dublin, Ireland.
Pseudocode No No pseudocode or algorithm blocks are explicitly presented in the paper. The paper contains mathematical equations and derivations but no structured algorithm formats.
Open Source Code Yes The code associated with this work can be found at https://github.com/deepmind/deepmind-research/dd_two_player_games.
Open Datasets Yes To understand the effect of DD on more complex adversarial games, we analyze GANs trained for image generation on the CIFAR10 dataset.
Dataset Splits No The paper mentions using the CIFAR10 dataset but does not explicitly provide details about the training, validation, and test splits (e.g., percentages or sample counts) within the main text.
Hardware Specification No The paper does not specify any hardware details (e.g., GPU models, CPU types, or memory) used for running the experiments.
Software Dependencies No The paper mentions optimizers like SGD and Adam, but it does not list any specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For SGD we use learning rates {5 10 2, 1 10 2, 5 10 3, 1 10 3} for each player; for Adam, we use learning rates {1 10 4, 2 10 4, 3 10 4, 4 10 4}, which have been widely used in the literature (Miyato et al., 2018). When comparing to Runge-Kutta (RK4) to assess the effect of DD we always use the same learning rates for both players.