Fast Instrument Learning with Faster Rates

Authors: Ziyu Wang, Yuhao Zhou, Jun Zhu

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Simulation studies demonstrate the competitive performance of our method. (Abstract) and We use a DNN as the black-box learner, and a RBF kernel for H, with bandwidth determined by marginal likelihood (72). We set N1 = N2 {500, 2500, 5000}. (Section 7)
Researcher Affiliation Collaboration Ziyu Wang, Yuhao Zhou, Jun Zhu Dept. of Comp. Sci. and Tech., BNRist Center, State Key Lab for Intell. Tech. & Sys., Institute for AI, Tsinghua-Bosch Joint Center for ML, Tsinghua University {wzy196,yuhaoz.cs}@gmail.com, dcszj@tsinghua.edu.cn
Pseudocode Yes Algorithm 1 Kernelized IV with learned instruments.
Open Source Code Yes Code available at https://github.com/meta-inf/fil.
Open Datasets Yes Our main simulation setup is adapted from [18, 19]; Appendix H.4 presents additional experiment on the demand dataset [6, 17]. and MNIST [62] or CIFAR-10 [63] image
Dataset Splits No The paper discusses validation statistics and sample sizes (n1, n2) but does not explicitly specify the percentages or counts for training, validation, or test splits. It implicitly uses standard splits for datasets like MNIST/CIFAR-10 but doesn't state them for the custom simulation setup.
Hardware Specification Yes All experiments are conducted on a single Nvidia RTX 3090 GPU, and we use a single core for experiments involving tree-based methods. (Appendix H.1)
Software Dependencies No The paper mentions using DNNs and kernels but does not specify the versions of software libraries (e.g., PyTorch, TensorFlow, scikit-learn) or programming languages used.
Experiment Setup Yes We use a DNN as the black-box learner, and a RBF kernel for H, with bandwidth determined by marginal likelihood (72). We set N1 = N2 {500, 2500, 5000}. (Section 7). DNN: Three fully connected layers with 128 neurons and ReLU activation. Regularization parameter 10 3. Training: 100 epochs with Adam optimizer and learning rate 0.001. (Appendix H.1.2)