Metric-Optimized Example Weights

Authors: Sen Zhao, Mahdi Milani Fard, Harikrishna Narasimhan, Maya Gupta

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We illustrate the performance of the proposed method on diverse public benchmark datasets and real-world applications. In this section, we illustrate the value of our proposal by comparing it to common strategies on a diverse set of example problems.
Researcher Affiliation Industry 1Google AI, 1600 Amphitheatre Parkway, Mountain View, CA 94043, USA. Correspondence to: Sen Zhao <senzhao@google.com>.
Pseudocode Yes Algorithm 1 Get optimal ˆα and ˆθ(ˆα) and Algorithm 2 Get Candidate αi+1
Open Source Code Yes The code on public datasets is available at the following Git Hub address: https://github.com/google-research/googleresearch/tree/master/moew.
Open Datasets Yes MNIST handwritten digit database (Le Cun & Cortes, 2010), wine reviews dataset from Kaggle (www.kaggle.com/zynicide/wine-reviews), Communities and Crime dataset from the UCI Machine Learning Repository (Dheeru & Karra Taniskidou, 2017)
Dataset Splits Yes training/validation/test split of sizes 55k/5k/10k respectively. (MNIST), training/validation/test split of sizes 85k/12k/24k respectively. (Wine Reviews), 994/500/500 training/validation/testing examples purely randomly. (Communities and Crime)
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models or exact server configurations used for experiments.
Software Dependencies No The paper mentions using 'Adam optimizer' but does not specify software dependencies with version numbers (e.g., TensorFlow, PyTorch versions).
Experiment Setup Yes Both the autoencoder and the main models were trained for 10k steps using Adam optimizer (Kingma & Ba, 2015) with learning rate 0.001. We used squared loss for numeric, hinge loss for binary, and cross-entropy loss for multiclass label/features. We sampled B K candidate α s in a d-dimensional ball of radius R using GP-BUCB with p = q = 68.3 and an RBF kernel, whose kernel width was set to be equal to R.