Unsupervised Learning under Latent Label Shift

Authors: Manley Roberts, Pranav Mani, Saurabh Garg, Zachary Lipton

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

Reproducibility Variable Result LLM Response
Research Type Experimental In semi-synthetic experiments, we adapt existing image classification benchmarks to the LLS setting, sampling without replacement to construct collections of label-shifted domains. and Table 1: Results on CIFAR-20. Each entry is produced with the averaged result of 3 different random seeds.
Researcher Affiliation Academia Manley Roberts Pranav Mani Saurabh Garg Zachary C. Lipton Carnegie Mellon University {manleyroberts,zlipton}@cmu.edu; {pmani, sgarg2}@cs.cmu.edu
Pseudocode Yes Algorithm 1 DDFA Training and Algorithm 2 DDFA Prediction
Open Source Code Yes Code is available at https://github.com/acmi-lab/Latent-Label-Shift-DDFA.
Open Datasets Yes First we examine standard multiclass image datasets CIFAR-10, CIFAR-20 [43], and Image Net-50 [19] containing images from 10, 20, and 50 classes respectively.
Dataset Splits Yes Algorithm 1 DDFA Training: 1: Split into train set T and validation set V and The algorithm uses train and validation data consisting of pairs of images and domain indices. We train Res Net50 [35] (with added dropout) on images xi with domain indices di as the label, choose best iteration by valid loss, pass all training and validation data through pf, and cluster pushforward predictions pfpxiq into m ě k clusters with Faiss K-Means [42].
Hardware Specification Yes We use one NVIDIA V100 GPU for all experiments.
Software Dependencies No The paper mentions several software components and libraries, such as 'Res Net50', 'Sim CLR', 'Mo Co', 'Faiss K-Means', and 'Sklearn implementation of standard NMF', but it does not specify any version numbers for these dependencies.
Experiment Setup Yes We train for 100 epochs with a learning rate of 1e-4, batch size of 256, and Adam optimizer. We use a learning rate of 1e-3, batch size of 256 for SCAN.