Learning from weak labelers as constraints

Authors: Vishwajeet Agrawal, Rattana Pukdee, Nina Balcan, Pradeep K Ravikumar

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, we demonstrate the superior performance and robustness of our method on a popular weak supervision benchmark. ... 5 EXPERIMENTAL EVALUATION ... Table 1: Average test accuracy and the corresponding standard error (over 5 random train-val-test split of the data) of our proposed algorithm and the baselines.
Researcher Affiliation Academia Vishwajeet Agrawal*, Rattana Pukdee*, Maria-Florina Balcan, Pradeep Ravikumar Carnegie Mellon University EMAIL
Pseudocode Yes We also provide a compact version of our algorithm in Algorithm 1 in Appendix C. ... Algorithm 1 Learning from weak labeler constraints
Open Source Code No The paper does not provide an explicit statement or link for the open-sourcing of the code for the methodology described.
Open Datasets Yes We show a comparison of our proposed method and baselines on 8 text classification datasets from the WRENCH benchmark Zhang et al. (2021).
Dataset Splits Yes Table 1: Average test accuracy and the corresponding standard error (over 5 random train-val-test split of the data) of our proposed algorithm and the baselines. ... We tune these hyperparameters on the validation set of size 100 for each dataset.
Hardware Specification No The paper does not provide specific hardware details used for running the experiments. It only mentions using a neural network and training parameters.
Software Dependencies No The paper mentions 'BERT text embeddings', 'Adam optimizer', 'scipy.linprog in Python' and 'cvxpy.CLARABEL convex solver' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes For all methods and datasets, we use a neural network with a single hidden layer and a hidden size of 16 on top of the pre-trained BERT text embeddings. The neural network is trained with a full batch gradient descent with an Adam optimizer with a learning rate in [0.001, 0.003, 0.01], weight decay in [0.001, 0.003, 0.01] and a number of epochs in range(1, 500, 5).