Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Prior Learning in Introspective VAEs
Authors: Ioannis Athanasiadis, Fredrik Lindsten, Michael Felsberg
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we perform a series of targeted experiments on a 2D density estimation benchmark and in an image generation setting comprised of the (F)-MNIST and CIFAR-10 datasets demonstrating the effect of prior learning in S-Intro VAE in generation and representation learning. [...] In this section we investigate the impact of learning the prior in S-Intro VAE. Our testbed consists of a 2D density estimation benchmark alongside three image datasets of varying complexity. [...] We evaluate the generation quality using the FID metric for samples generated from sampling from the prior and the aggregated posterior denoted as FID(GEN) and FID(REC) respectively. [...] The quality of the representations learned by the encoder was evaluated by fitting a linear SVM, similar to Kviman et al. (2023), using 2K-SVM and 10K-SVM iterations as well as utilizing a k-nearest neighbor classifier (k-NN) using 5-NN or 100-NN (Caron et al., 2021). |
| Researcher Affiliation | Academia | Ioannis Athanasiadis EMAIL Department of Electrical Engineering Linköping University Fredrik Lindsten EMAIL Department of Computer and Information Science Linköping University Michael Felsberg EMAIL Department of Electrical Engineering Linköping University |
| Pseudocode | Yes | In this section, we outline the implementation choices as well as the motivation behind them enabling prior learning in S-Intro VAE in a prior decoder cooperation manner. Pseudo-code for the prior learning in SIntro VAE is provided in Algorithm 1. |
| Open Source Code | No | The paper does not contain any explicit statement about making the code available for the work described, nor does it provide a link to a code repository. |
| Open Datasets | Yes | Finally, we perform a series of targeted experiments on a 2D density estimation benchmark and in an image generation setting comprised of the (F)-MNIST and CIFAR-10 datasets demonstrating the effect of prior learning in S-Intro VAE in generation and representation learning. [...] Our testbed consists of a 2D density estimation benchmark alongside three image datasets of varying complexity. |
| Dataset Splits | No | The paper mentions using well-known datasets like (F)-MNIST and CIFAR-10, which typically have standard splits. It also mentions evaluating representations by fitting a linear SVM using "2K-SVM and 10K-SVM iterations" and k-NN classifiers. However, it does not explicitly provide specific dataset split percentages, sample counts for training/validation/test sets for the main model training, or citations to predefined splits for the experimental setup. |
| Hardware Specification | No | The computational resources were provided by the National Academic Infrastructure for Supercomputing in Sweden (NAISS) at C3SE, partially funded by the Swedish Research Council through grant agreement no. 2022-06725. This statement mentions the resource provider but does not include specific hardware details such as GPU/CPU models, memory, or processor types. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers, such as programming language versions or library versions (e.g., Python 3.x, PyTorch 1.x). |
| Experiment Setup | Yes | In line with Daniel & Tamar (2021), we identified the optimal hyperparameters (i.e., βrec, βKL and βneg) by performing an extensive grid-search while we used α = 2 and γ = 1. [...] For our all our experiments we used K = 10. [...] Algorithm 1 Prior Learning in S-Intro VAE (Daniel & Tamar, 2021). The red-highlighted segments indicate the parts that differ from the standard Gaussian S-Intro VAE. The Lrec and the LKL refer to the reconstruction loss and the KL divergence between the posterior and the prior target respectively, whereas the Lsg KL is a modified KL divergence that applies the stop-gradient sg operator on the prior as target. Require: βrec, βKL, βneg, γ, η, rentropy, K |