BR-SNIS: Bias Reduced Self-Normalized Importance Sampling

Authors: Gabriel Cardoso, Sergey Samsonov, Achille Thin, Eric Moulines, Jimmy Olsson

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we propose a new method, BR-SNIS, whose complexity is essentially the same as that of SNIS and which significantly reduces bias without increasing the variance. This method is a wrapper in the sense that it uses the same proposal samples and importance weights as SNIS, but makes clever use of iterated sampling importance resampling (i-SIR) to form a bias-reduced version of the estimator. We furnish the proposed algorithm with rigorous theoretical results, including new bias, variance and high-probability bounds, and these are illustrated by numerical examples.
Researcher Affiliation Academia Gabriel Cardoso Centre de Mathématiques appliquées, Ecole polytechnique, IHU Liryc, fondation Bordeaux Université, Univ Bordeaux, CRCTB U4045, INSERM, gabriel.victorino-cardoso@polytechnique.edu Sergey Samsonov HSE University Achille Thin UMR MIA, Agro Paris Tech, Eric Moulines Centre de Mathématiques appliquées, Ecole polytechnique, Jimmy Olsson Department of Mathematics, KTH Royal Institute of Technology.
Pseudocode No The paper describes algorithms (i-SIR, BR-SNIS) in text, but it does not provide any explicitly labeled "Pseudocode" block or "Algorithm" figure.
Open Source Code Yes The code used for this experiment is available at 1. https://github.com/gabrielvc/br_snis
Open Datasets Yes For numerical illustration, we use the heart failure clinical records (d = 13, T = 299), breast cancer detection (d = 30, T = 569), and Covertype (d = 55, T = 4 104) datasets from the UCI machine learning repository. ... As an illustration, we train the model using the binarized MNIST dataset [41]
Dataset Splits No The paper mentions the "MNIST validation set" in Table 2, but it does not specify the exact splits (percentages or counts) used for training, validation, or testing for any of the datasets.
Hardware Specification No The paper does not mention any specific GPU models, CPU models, or cloud computing instance types used for the experiments. It refers to "computational resources" generally.
Software Dependencies No The paper mentions using the "Adam optimizer [20]" but does not specify its version number or any other software dependencies with their versions (e.g., Python, PyTorch, TensorFlow versions).
Experiment Setup Yes For the algorithm BR-SNIS, we used N {129, 513}, k0 = kmax 1 and kmax = M/(N 1) bootstrap rounds. ... All models are run for 100 epochs, using the Adam optimizer [20] and a learning rate of 10 4. ... As a prior, we use a Gaussian distribution N(0, τ 2I) with τ 2 = 5 10 2.