Resource Constrained Structured Prediction

Authors: Tolga Bolukbasi, Kai-Wei Chang, Joseph Wang, Venkatesh Saligrama

AAAI 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our proposed adaptive system on two structured prediction tasks, optical character recognition and dependency parsing and show significant reduction in the feature costs without degrading accuracy.
Researcher Affiliation Academia Tolga Bolukbasi, Kai-Wei Chang,+ Joseph Wang, Venkatesh Saligrama Boston University, Boston, MA +University of Virginia, Chancellorsville, VA tolgab@bu.edu, kw@kwchang.net, joewang@bu.edu, and srv@bu.edu
Pseudocode Yes Algorithm 1 Anytime Policy Learning
Open Source Code No The paper provides a link to an arXiv preprint (https://arxiv.org/abs/1602.08761) for 'Implementation details and proofs', but does not explicitly state that the source code for their methodology is available there or elsewhere.
Open Datasets Yes We tested our algorithm on a sequence-label problem, the OCR dataset (Taskar, Guestrin, and Koller 2003)...
Dataset Splits No For the OCR dataset, the paper states, 'We split the data such that 90% is used for training and 10% is used for test,' but does not explicitly mention a validation split. For the Penn Treebank dataset, no specific training, validation, or test splits are provided.
Hardware Specification No The paper provides performance metrics in terms of time (e.g., '165 μs per word', '275 μs per word', '75 μs per word') and notes that other work uses 'different features, policy settings, and hardware', but it does not specify the particular hardware (CPU, GPU models, etc.) used for its own experiments.
Software Dependencies No The paper mentions that 'All algorithms are implemented based on the graph-based dependency parser (Mc Donald et al. 2005) in Illinois-SL library (Chang et al. 2015)', but it does not provide specific version numbers for this library or any other software dependencies such as programming languages or frameworks.
Experiment Setup No The paper mentions some setup details such as 'sweeping λ in Equation 1', using 'simple features' and 'a weighted SVM' for the policy, and specifying feature types like 'raw pixel values and HOG features'. However, it does not provide concrete hyperparameter values (e.g., learning rate, batch size, epochs, specific SVM parameters) or system-level training settings.