Bootstrapping the Error of Oja's Algorithm

Authors: Robert Lunde, Purnamrita Sarkar, Rachel Ward

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

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Experimental validation of the online multiplier bootstrap
Researcher Affiliation Academia Robert Lunde University of Michigan rlunde@umich.edu Purnamrita Sarkar University of Texas at Austin purna.sarkar@austin.utexas.edu Rachel Ward University of Texas at Austin rward@math.utexas.edu
Pseudocode Yes Algorithm 1: Bootstrap for Oja s algorithm
Open Source Code No The paper does not provide an explicit statement or link for open-source code for the described methodology.
Open Datasets No We draw Zij IID Uniform(-√3, √3), for i = 1, . . . , n and j = 1, . . . d. Consider a PSD matrix Kij = exp(-|i - j|c) with c = 0.01. We create a covariance matrix such that Σij = K(i, j)σiσj. We consider σi = 5i-β for β = 0.2 and β = 1. Now we transform the data to introduce dependence by letting Xi = Σ1/2Zi. By construction, we have that E[Xi Xi^T] = Σ for all 1 <= i <= n.
Dataset Splits No The paper describes generating synthetic datasets but does not specify any explicit training, validation, or test data splits.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers.
Experiment Setup Yes We use ηn = log n. In Figure 1, we see that for β = 0.2 (see (A) and (B)), where the variance decay is slow and therefore the error bounds of the residual terms are expected to be large, the quality of approximation is poorer compared to (C) and (D), where β = 1. We draw g N(0, Id) Create unit vector u0 g/ g Initialize ˆv1, v (1) 1 , . . . , v (m) 1 u0 for t=2,..., n do Update ˆv1 ˆv1 + η(XT t ˆv1)ˆv1 Normalize ˆv1 to have unit norm; for i=1:m do Draw Wi N(0, 1/2); and We draw Zij IID Uniform(-√3, √3), for i = 1, . . . , n and j = 1, . . . d. and for n = 1000, d = 500 in (A) and (C) and for n = 10, 000, d = 500 in (B) and (D).