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 [1].
Injective flows for star-like manifolds
Authors: Marcello Negri, Jonathan Aellen, Volker Roth
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We showcase the generality of our approach with two very distinct applications in Bayesian modeling. First, we introduce a novel Objective Bayesian approach to penalized likelihood methods. [...] In Section 5.1 we show empirically that the proposed approach provides a significant speedup compared to the explicit computation of the Jacobian determinant. We also show that in variational inference the exact Jacobian determinant is crucial for training while the approximation commonly employed in injective flows results in poor reconstruction. [...] In Figure 3 we compare the ground truth (log) density with that obtained with the two models. [...] In Figure 4 we show that the modeled densities perfectly match the ground truth in all cases. |
| Researcher Affiliation | Academia | Marcello Massimo Negri & Jonathan Aellen & Volker Roth Department of Mathematics and Computer Science, University of Basel EMAIL |
| Pseudocode | No | The paper describes the methodology in prose and mathematical equations. While it provides details on the implementation steps and architecture (e.g., in Appendix A.5 and A.6.1), it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks, nor does it present structured steps in a code-like format. |
| Open Source Code | Yes | For the implementation we rely on the (conditional) normalizing flow library Flow Conductor1, which was introduced in Negri et al. (2023) and Arend Torres et al. (2024). 1https://github.com/Fabricio Arend Torres/Flow Conductor |
| Open Datasets | Yes | We select a portfolio with 10 stocks over a period of 200 time steps from the dataset in Tu and Li (2024). [Tu and Li (2024) is listed in references as: Xueyong Tu and Bin Li. Robust portfolio selection with smart return prediction. Economic Modelling, 135(C), 2024.] |
| Dataset Splits | No | The paper mentions creating a synthetic regression dataset by sampling from distributions and selecting a portfolio from a cited dataset, but it does not specify any training, validation, or test splits, nor percentages or specific counts for data partitioning. |
| Hardware Specification | Yes | Notably, all trained flows converged in a matter of minutes on a standard consumer-grade GPU (RTX2080Ti in our specific case). |
| Software Dependencies | No | The paper mentions several software components and libraries like Flow Conductor, circular spline layers (referencing Rezende et al., 2020 and Stimper et al., 2023), neural spline layers (Durkan et al., 2019), and the Adam optimizer (Kingma and Ba, 2017). However, it does not provide specific version numbers for these components or for the programming language used. |
| Experiment Setup | Yes | One is for standard NFs that we use for the subjective penalized likelihood regression problem. The other architecture is the injective flow that is used for the objective Bayes version of the regression problem and the portfolio diversification application. [...] It consists of a normal distribution as base distribution. Then we use 5 blocks of permutation transformation, a sum of Sigmoids layer (Negri et al., 2023) and an activation norm. The sum of Sigmoid layer consists each of 30 individual Sigmoid functions in three blocks. [...] The base distribution is either the probabilistic simplex or the complete β 1 = 1 depending on the application. We follow this with again 5 layers of the circular bijective layers (Rezende et al., 2020), each consisting of three blocks with 8 bins. [...] We optimize the reverse KL divergence using Adam (Kingma and Ba, 2017) as optimizer with default parameters. [...] The synthetic regression dataset is created by sampling X from a 5 dimensional Wishart distribution W5(7, I). The response variable y is then created by X β + ϵ where β is standard normal distributed and ϵ is normal distributed with zero mean and a standard deviation of 4.0. [...] Our sampler uses 1000 chains each with 100 000 samples. |