Not-So-Random Features

Authors: Brian Bullins, Cyril Zhang, Yi Zhang

ICLR 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Evaluations on synthetic and real-world datasets demonstrate scalability and consistent improvements over related random features-based methods. Finally, we exhibit experiments on synthetic and benchmark datasets, demonstrating consistent improvements over related random features-based kernel methods. In this section, we highlight the most important and illustrative parts of our experimental results.
Researcher Affiliation Academia Brian Bullins Cyril Zhang Yi Zhang Department of Computer Science Princeton University Princeton, NJ 08544, USA {bbullins, cyril.zhang, y.zhang}@cs.princeton.edu
Pseudocode Yes Algorithm 1 Langevin dynamics for kernel alignment. Algorithm 2 No-regret learning dynamics for SVM margin maximization.
Open Source Code Yes The code can be found at github.com/yz-ignescent/Not-So-Random-Features.
Open Datasets Yes Challenging label pairs are chosen from the MNIST (Le Cun et al., 1998) and CIFAR-10 (Krizhevsky, 2009) datasets;
Dataset Splits No For synthetic data: 'ntrain = 2000 and ntest = 50000'. For MNIST/CIFAR-10: 'each task consists of 10000 training and 2000 test examples'. No explicit mention or details of a validation set split.
Hardware Specification No The paper mentions 'an efficient GPU implementation' but does not provide specific hardware details like GPU or CPU models, memory, or detailed cloud/cluster configurations used for experiments.
Software Dependencies No The paper does not specify any software dependencies with version numbers.
Experiment Setup Yes Throughout all experiments presented, we use hinge-loss SVM classifiers with C = 1. With regard to Langevin diffusion (Algorithm 1), we observe that the best samples arise from using high temperatures and Gaussian parallel initialization. For the latter, a rule-of-thumb is to initialize 500 parallel copies of Langevin dynamics... Empirically, running Langevin dynamics for 100 steps suffices to locate a reasonably good peak. The step size of online gradient ascent is set to balance between being conservative and promoting diverse samples;