Learning Dependency Structures for Weak Supervision Models
Authors: Paroma Varma, Frederic Sala, Ann He, Alexander Ratner, Christopher Re
ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our structure learning method on real-world applications ranging from medical image classification to relation extraction over text. We compare our performance to several common weak supervision baselines: an unweighted majority vote of the weak supervision source labels, a generative modeling approach that assumes independence among weak supervision sources (Ratner et al., 2016), and a generative model using dependency structure learned with an existing structure learning approach for weak supervision (Bach et al., 2017). We report performance of the discriminative model trained on labels from these generative models in Table 2. Finally, we run simulations to explore the performance of our method under two conditions from Section 3. |
| Researcher Affiliation | Academia | 1Department of Electrical Engineering, Stanford University 2Department of Computer Science, Stanford University. |
| Pseudocode | Yes | Algorithm 1 Weak Supervision Structure Learning |
| Open Source Code | No | The paper does not provide any statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | The Bone Tumor task is to classify tumors in X-rays as aggressive or non-aggressive (Varma et al., 2017b). The CDR task is to detect relations among chemicals and disease mentions in Pub Med abstracts (Bach et al., 2017; Wei et al., 2015). The IMDb task is to classify plot summaries as describing action or romantic movies (Varma et al., 2017c). The MS-COCO task is to classify images as containing a person (Varma et al., 2017a). |
| Dataset Splits | No | The paper mentions using weak supervision sources to efficiently assign training labels and training a downstream discriminative model, but it does not specify any training/validation/test dataset splits, nor does it reference predefined splits with citations for reproducibility. |
| Hardware Specification | No | The paper does not explicitly describe the hardware used for running its experiments, such as specific GPU or CPU models. |
| Software Dependencies | No | The paper mentions using "standard convex solvers" but does not provide specific software names with version numbers required to replicate the experiments. |
| Experiment Setup | No | The paper describes the tasks and the general types of discriminative models used (logistic regression, LSTM, GoogLeNet). However, it does not provide specific experimental setup details such as hyperparameter values (e.g., learning rates, batch sizes, number of epochs) or other system-level training settings. |