On Recovering from Modeling Errors Using Testing Bayesian Networks
Authors: Haiying Huang, Adnan Darwiche
ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present empirical results in Section 6 and close with some concluding remarks in Section 7.We now empirically evaluate the techniques we proposed in Sections 4 and 5. We consider two tasks: model compression and supervised learning. Both tasks try to recover a function Pr(Q|E) generated by a true BN (G, Θ) but using an incomplete BN structure G that may be the result of a modeling error. The recovery is accomplished using a TBN with structure G . In the first task, we handcraft TBN parameters (CPTs and intervals) based on the true BN and the policy of Section 5 (see Equation 3). The goal is to assess how well can a TBN reach the promise of full recovery from modeling errors using a finite number of intervals. In the second task, we assess whether learning TBN parameters, instead of handcrafting them, can potentially reach the promise of full recovery by comparing the learned function with the handcrafted and true ones. |
| Researcher Affiliation | Academia | 1Computer Science Department, University of California, Los Angeles, USA. Correspondence to: Haiying Huang <hhaiying@ucla.edu>. |
| Pseudocode | No | The paper describes methods and processes but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper mentions using the PYTAC system, but does not state that the authors are releasing their own code for the methodology described in this paper. |
| Open Datasets | No | For this experiment, we randomly generate polytree BNs and queries Pr(Q|E). We simulate a modeling error by abstracting some nodes of the polytree to yield an incomplete structure G . We then parameterize G according to the soft policy (Equation 1) and average policy (Equation 2) to yield two BNs that we call BN SOFT and BN AVG. Next, we generate labeled datasets E, Q from the BNs (true models), each including 100 n examples where n is the number of distinct evidence instantiations. |
| Dataset Splits | No | Next, we generate labeled datasets E, Q from the BNs (true models), each including 100 n examples where n is the number of distinct evidence instantiations. No explicit training/validation/test splits are mentioned. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware (e.g., GPU models, CPU types, memory) used to run the experiments. |
| Software Dependencies | No | The paper mentions the 'PYTAC system' and generally refers to 'gradient descent' and 'tensor graphs' but does not specify any software names with version numbers for reproducibility. |
| Experiment Setup | No | The paper does not provide specific experimental setup details such as hyperparameter values (e.g., learning rate, batch size, number of epochs, optimizer settings) used for training. |