Rejection via Learning Density Ratios

Authors: Alexander Soen, Hisham Husain, Philip Schulz, Vu Nguyen

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our framework is tested empirically over clean and noisy datasets.
Researcher Affiliation Collaboration Alexander Soen Amazon The Australian National University alexander.soen@anu.edu.au Hisham Husain hisham.husain@protonmail.com Philip Schulz Amazon phschulz@amazon.com Vu Nguyen Amazon vutngn@amazon.com
Pseudocode Yes Algorithm 1 Density-Ratio Rejection
Open Source Code Yes Our rejector s code public at: https://github.com/alexandersoen/density-ratio-rejection.
Open Datasets Yes We consider 6 multiclass classification datasets. For tabular datasets, we consider the gas drift dataset [68] and the human activity recognition (HAR) dataset [5]...the MNIST image dataset [46] (10 classes)...CIFAR-10 [44] (10 classes); and Org MNIST / Organ SMNIST (11 classes) and Oct MNIST (4 classes) from the Med MNIST collection [69, 70]. All datasets we consider are in the public domain, e.g., UCI [6].
Dataset Splits Yes In our case, we have a fixed ρ which allows easy tuning of τ given a validation dataset, similar to other confidence based rejection approaches, e.g., tuning a threshold for the margin of a classifier [7]. All evaluation uses 5-fold cross validation.
Hardware Specification Yes All implementation use Py Torch and training was done on a p3.2xlarge AWS instance.
Software Dependencies No All implementation use Py Torch and training was done on a p3.2xlarge AWS instance. No specific version numbers for software dependencies are provided.
Experiment Setup Yes For our tests, we fix λ = 1. Throughout our evaluation, we assume that a neural network (NN) model without rejection is accessible for all (applicable) approaches...For our density ratio rejectors, we utilize the log-loss, practical considerations in Section 4.3, and Algorithm 1...Each is trained with 50 equidistant costs c, τ [0, 0.5), except on Oct MNIST which uses 10 equidistant costs...All training utilizes the Adam [41] optimizer. (And specific network architectures, batch sizes, epochs, learning rates are detailed in Appendix Q.I).