A Quantile-based Approach for Hyperparameter Transfer Learning

Authors: David Salinas, Huibin Shen, Valerio Perrone

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate significant improvements over state-of-the-art methods for both hyperparameter optimization and neural architecture search.
Researcher Affiliation Industry 1NAVER LABS Europe (work started while being at Amazon) 2Amazon Web Services. Correspondence to: David Salinas <david.salinas@naverlabs.com>, Huibin Shen <huibishe@amazon.com>, Valerio Perrone <vperrone@amazon.com>.
Pseudocode Yes Pseudo-code is given in Algorithm 1. ... Pseudo-code is given in Algorithm 2.
Open Source Code No The paper does not provide an explicit statement or link for the open-sourcing of the methodology's code.
Open Datasets Yes We consider three algorithms in the HPO context: XGBoost (Chen & Guestrin, 2016), a 2-layer feed-forward neural network (FCNET) (Klein & Hutter, 2019), and the RNN-based time series prediction model proposed in Salinas et al. (2017) (Deep AR). ... We also run experiments on NAS-Bench-201 (Dong & Yang, 2020). ... The list of the datasets is in the appendix.
Dataset Splits Yes We compute tabular evaluations (log) uniformly beforehand on multiple datasets to compare methods with sufficiently many random repetitions. ... The transfer learning capabilities of each method are evaluated in a leave-one-task-out setting: one dataset is sequentially left out to assess how much transfer can be achieved from the other datasets, and overall results are aggregated.
Hardware Specification Yes We run each experiment with 30 random seeds on AWS batch with m4.xlarge instances.
Software Dependencies No The paper mentions using 'GPareto (Binois & Picheny, 2019)' but does not specify its version number or any other software dependencies with explicit version details.
Experiment Setup Yes The MLP hwh(x) used to regress µθ and σθ has 3 layers with 50 nodes, a dropout rate of 0.1 after each hidden layer and relu activation functions. The learning rate is set to 0.01, and ADAM is run over 1000 gradient updates three times, lowering the learning rate by 5 each time with a batch size of 64.