Easy Learning from Label Proportions

Authors: Róbert Busa-Fekete, Heejin Choi, Travis Dick, Claudio Gentile, Andres Munoz Medina

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we empirically evaluate EASYLLP, PROPMATCH, and two baseline methods to characterize how their performance depends on the bag size for a range of different learning tasks and underlying learning models. ... Results. Figure 2 depicts the accuracy achieved by each method on a selection of datasets and models for a range of bag sizes.
Researcher Affiliation Industry Robert Busa-Fekete Google Research busarobi@google.com; Heejin Choi Coupang Inc hechoi53@coupang.com; Travis Dick Google Research tdick@google.com; Claudio Gentile Google Research cgentilek@google.com; Andres Munoz Medina Google Research ammmedina@google.com
Pseudocode Yes We study a version of projected SGD that picks one example per bag and uses the soft-label corrected gradient estimates (pseudocode is given in in Algorithm 1 in the Appendix).
Open Source Code No The paper does not provide an explicit statement or link for the open-source code of their methodology.
Open Datasets Yes We carry out experiments on four (binary classification) datasets: Binarized versions of MNIST [13] and CIFAR-10 [12], as well as the Higgs [3] and UCI adult datasets [11].
Dataset Splits No The paper mentions tuning learning rates and hyperparameters, which implies the use of a validation set, but does not explicitly provide specific details about training/test/validation dataset splits (e.g., percentages or sample counts) for reproducibility.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU or CPU models, memory specifications) used for running the experiments.
Software Dependencies No The paper mentions using 'Adam [10] optimizer' and refers to frameworks like 'Tensorflow, JAX and Py Torch' but does not specify exact version numbers for any software dependencies.
Experiment Setup Yes To tune the learning rate for each method, we report the highest accuracy achieved for learning rates in {0.00001, 0.0001, 0.0005, 0.001, 0.005, 0.01, 0.05}. ... In all cases we use the Adam [10] optimizer, binary crossentropy loss, minibatches of size 512, and 20 training passes through the data. Finally, for the two image datasets, we decay the learning rate after 40%, 60%, 80%, and 90% of the training passes by factors 10, 100, 1000, and 5000, respectively.