ADAVI: Automatic Dual Amortized Variational Inference Applied To Pyramidal Bayesian Models
Authors: Louis Rouillard, Demian Wassermann
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the capability and scalability of our method on simulated data, as well as a challenging high-dimensional brain parcellation experiment. |
| Researcher Affiliation | Academia | Louis Rouillard Universit e Paris-Saclay, Inria, CEA Palaiseau, 91120, France louis.rouillard-odera@inria.fr Demian Wassermann Universit e Paris-Saclay, Inria, CEA Palaiseau, 91120, France demian.wassermann@inria.fr |
| Pseudocode | No | The paper does not include a clearly labeled 'Pseudocode' or 'Algorithm' block or figure. |
| Open Source Code | Yes | As part of our submission we release the code associated to our experiments. |
| Open Datasets | Yes | Data is extracted from the Human Connectome Project dataset (Van Essen et al., 2012). |
| Dataset Splits | Yes | To measure the convergence, we compute analytically the KL divergence between the variational posterior and the analytical one (every distribution being a Gaussian), summed for every distribution, and averaged over a validation dataset of size 2000. |
| Hardware Specification | Yes | All experiments were performed on a computational cluster with 16 Intel(R) Xeon(R) CPU E5-2660 v2 @ 2.20GHz (256Mb RAM), 16 AMD EPYC 7742 64-Core Processor (512Mb RAM) CPUs and 1 NVIDIA Quadro RTX 6000 (22Gb), 1 Tesla V100 (32Gb) GPUs. |
| Software Dependencies | No | All methods were implemented in Python. We implemented most methods using Tensorflow Probability (Dillon et al., 2017), and SBI methods using the SBI Python library (Tejero-Cantero et al., 2020). The paper mentions software but lacks specific version numbers for the libraries. |
| Experiment Setup | Yes | ADAVI: NF with 1 Affine block with triangular scale, followed by 1 MAF with [32, 32, 32] units. HE with embedding size 8, 2 modules with 2 ISABs (2 heads, 8 inducing points), 1 PMA (seed size 1), 1 SAB and 1 linear unit each. Minibatch size 32, 32 theta draws per X point (see appendix C.5), Adam (10 3), 40 epochs using a reverse KL loss. |