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; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Tractable Approximate Gaussian Inference for Bayesian Neural Networks
Authors: James-A. Goulet, Luong Ha Nguyen, Saeid Amiri
JMLR 2021 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, Section 4 validates the performance of the approach on benchmark regression problems and on the MNIST classification problem. In this section, we perform experiments using the TAGI method for a 1D toy problem, for a set of benchmark regression problems, and for the MNIST classification dataset. |
| Researcher Affiliation | Academia | James-A. Goulet EMAIL Luong Ha Nguyen EMAIL Saeid Amiri Department of Civil Engineering, Polytechnique Montréal, Montréal, Canada |
| Pseudocode | No | The paper describes the methods using mathematical formulations and textual explanations, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks or figures. |
| Open Source Code | No | The paper states, "The timing details reported in Appendix E show that the current TAGI s implementation has a computational time that is more than four times faster than PBP and MC-Dropout (Gal and Ghahramani, 2016)." However, this refers to an internal implementation and there is no explicit statement or link indicating that the source code for the methodology is publicly available. |
| Open Datasets | Yes | Finally, Section 4 validates the performance of the approach on benchmark regression problems and on the MNIST classification problem. ... We apply TAGI to the MNIST classification problem (Le Cun et al., 1998) consisting of D = 70 000 (28 28) greyscale images for K = 10 classes (60 000 training and 10 000 test). |
| Dataset Splits | Yes | The optimal number of epochs is identified from a validation set DV consisting of 20 additional points. ... The initial value for the observation error s standard deviation is set to σV = 1, and this value is optimized using a 5-folds cross-validation setup. ... The optimal number of epochs E is identified using an early-stop procedure evaluated on the validation set. ... randomly selected validation set corresponding to 5% of the training set. |
| Hardware Specification | No | All these experiments were conducted using CPU. This statement is too general and does not provide specific details about the CPU model, other hardware components, or computing environment used for the experiments. |
| Software Dependencies | No | The paper does not explicitly mention any specific software libraries, frameworks, or tools along with their version numbers that were used to implement or run the experiments. |
| Experiment Setup | Yes | The optimal number of epochs is identified from a validation set DV consisting of 20 additional points. The prior covariance for bias is initialized to Σ0 B = 0.01 I, and for weights Σ0 W , by using the Xaviers s approach (Glorot and Bengio, 2010). The prior mean vector is randomly sampled from µ0 θ N(0, Σ0 θ) in order to break the initial symmetry in the network, as starting with zero expected values for weights adversely affects learning. ... For all cases, the data is normalized, the activation function is a Re LU, and the batch size is B = 10; The prior covariance for biases is initialized to Σ0 B = 0.01 I, and for weights Σ0 W , by using the Xaviers s approach (Glorot and Bengio, 2010); The initial value for the observation error s standard deviation is set to σV = 1, and this value is optimized using a 5-folds cross-validation setup. ... Each network is evaluated for σV = {0.1, 0.2, 0.3, 0.4} and the optimal value for σV is selected using a randomly selected validation set corresponding to 5% of the training set. |