Gradient Boosted Normalizing Flows
Authors: Robert Giaquinto, Arindam Banerjee
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of this technique for density estimation and, by coupling GBNF with a variational autoencoder, generative modeling of images. Our results show that GBNFs outperform their non-boosted analog, and, in some cases, produce better results with smaller, simpler flows. 5 Experiments: To evaluate GBNF, we highlight results on two toy problems, density estimation on real data, and boosted flows within a VAE for generative modeling of images. We boost coupling flows [20, 46] parameterized by feed-forward networks with Tan H activations and a single hidden layer. |
| Researcher Affiliation | Academia | Robert Giaquinto Arindam Banerjee Department of Computer Science & Engineering University of Minnesota, Twin Cities Minneapolis, MN 55455, USA |
| Pseudocode | No | The paper does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement or link to open-source code for the described methodology. |
| Open Datasets | Yes | We report density estimation results on the POWER, GAS, HEPMASS, and MINIBOONE datasets from the UCI machine learning repository [23], as well as the BSDS300 dataset [55]. We compare our model on the same image datasets as those used in van den Berg et al. [78]: Freyfaces, Caltech 101 Silhouettes [54], Omniglot [49], and statically binarized MNIST [50]. |
| Dataset Splits | No | The paper refers to using datasets and evaluating performance (e.g., test loss in Figure 2, Table 1 and 2 reporting test set metrics), but it does not specify explicit training/validation/test splits with percentages or sample counts. It mentions |
| Hardware Specification | No | The paper does not provide specific hardware details (e.g., GPU/CPU models, memory specifications) used for running its experiments. It only mentions general computing support from the University of Minnesota Supercomputing Institute (MSI) in the Acknowledgements. |
| Software Dependencies | No | The paper mentions using |
| Experiment Setup | Yes | In the toy experiments flows are trained for 25k iterations using the Adam optimizer [45]. For all other experiments details on the datasets and hyperparameters can be found in Appendix A. In our experiments that augment the VAE with a GBF-based posterior, we find good results setting the regularization λ = 1.0. In the density estimation experiments, better results are often achieved with λ near 0.8. Further, we anneal the KL [6, 37, 72] in (14) cyclically [31], with restarts corresponding to the introduction of new boosting components |