Let Go of Your Labels with Unsupervised Transfer

Authors: Artyom Gadetsky, Yulun Jiang, Maria Brbic

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

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate TURTLE on a diverse benchmark suite of 26 datasets and show that it achieves new state-of-the-art unsupervised performance.
Researcher Affiliation Academia 1EPFL, Lausanne, Switzerland.
Pseudocode Yes The pseudocode of TURTLE is provided in Algorithm B1 with implementation details in Appendix B.3.
Open Source Code Yes Code is publicly available at https://github.com/mlbio-epfl/turtle.
Open Datasets Yes We study the performance of TURTLE on the extensive benchmark of 26 vision datasets (Radford et al., 2021). The detailed description of each dataset is provided in Appendix B.1.
Dataset Splits Yes We use 10-fold cross-validation to select the best task.
Hardware Specification Yes The selection process takes a few minutes on small datasets, and around 8 hours on Image Net, with a single NVIDIA A100 GPU.
Software Dependencies No We use the cu ML.Logistic Regression (Raschka et al., 2020) for linear probe evaluation in our paper 3. The cu ML library allows for GPU acceleration... ADAM (Kingma & Ba, 2015) optimizer is used for the training of both linear classifier and task encoder.
Experiment Setup Yes We use 10000 as the default batch-size. We update the linear classifier for M = 10 steps at each iteration and train the task encoder for T = 6000 iterations in total. If not specifically mentioned, we set the entropy regularization parameter γ = 10 for all experiments. We do a grid search over 5 different learning rates for both task encoder and linear classifier with η {0.01, 0.005, 0.001, 0.0005, 0.0001}, respectively.