Deep Variational Implicit Processes
Authors: Luis A. Ortega, Simon Rodriguez Santana, Daniel Hernández-Lobato
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We support these claims via extensive regression and classification experiments. We also evaluate DVIP on large datasets with up to several million data instances to illustrate its good scalability and performance. |
| Researcher Affiliation | Academia | 1Universidad Autónoma de Madrid 2ICMAT-CSIC {luis.ortega,daniel.hernandez}@uam.es, simon.rodriguez@icmat.es |
| Pseudocode | No | No pseudocode or clearly labeled algorithm block was found in the paper. |
| Open Source Code | Yes | An efficient Py Torch implementation of DVIP is found in the supplementary material. |
| Open Datasets | Yes | We compare each method on 8 regression datasets from the UCI Repository (Dua and Graff, 2017). ... We carry out experiments on the CO2 time-series dataset (https:// scrippsco2.ucsd.edu). ... First, the Year dataset (UCI) with 515, 345 instances and 90 features, where the original train/test splits are used. Second, the US flight delay (Airline) dataset (Dutordoir et al., 2020; Hensman et al., 2017)... Lastly, we consider data recorded on January, 2015 from the Taxi dataset (Salimbeni and Deisenroth, 2017). ... Rectangles (Salimbeni and Deisenroth, 2017) and the multi-class dataset MNIST (Deng, 2012). ... two massive binary datasets: SUSY and HIGGS, with 5.5 million and 10 million instances, respectively. These datasets contain Monte Carlo physics simulations to detect the presence of the Higgs boson and super-symmetry (Baldi et al., 2014). |
| Dataset Splits | No | The paper specifies train/test splits for various datasets (e.g., '20 different train / test splits of the data with 10% test size', 'original train/test splits are used', 'first 700, 000 instances for training and the next 100, 000 for testing'). However, it does not explicitly state a separate 'validation' split from the dataset partitioning. |
| Hardware Specification | No | The paper states, 'Authors gratefully acknowledge the use of the facilities of Centro de Computacion Cientifica (CCC) at Universidad Autónoma de Madrid,' but does not provide specific details on the CPU, GPU, or other hardware used for the experiments. |
| Software Dependencies | No | The paper mentions 'An efficient Py Torch implementation of DVIP is found in the supplementary material,' indicating the use of PyTorch, but it does not specify the version number of PyTorch or any other software dependencies. |
| Experiment Setup | Yes | We use S = 20 and a BNN as the IP prior for each unit. These BNNs have 2 layers of 10 units each with tanh activations, as in Ma et al. (2019). ... all models are trained for 100, 000 iterations. ... We set R to 100 for testing and to 1 for training, respectively. ... For these two datasets, results are averaged over 10 different random seed initializations. ... Here, we trained each method for 500, 000 iterations. ... In Rectangles, 20, 000 iterations are enough to ensure convergence. ... In the first layer we employ a convolutional NN (CNN) prior with two layers of 4 and 8 channels respectively. No input propagation is used in the first layer. ... train for 500, 000 iterations. ... Appendix C has all the details about the experimental settings considered for each method. |