Bayesian Model Selection, the Marginal Likelihood, and Generalization

Authors: Sanae Lotfi, Pavel Izmailov, Gregory Benton, Micah Goldblum, Andrew Gordon Wilson

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We investigate the correlation between the log marginal likelihood (LML) and generalization in the context of image classification using the CIFAR-10 and CIFAR-100 datasets.
Researcher Affiliation Academia Sanae Lotfi 1 Pavel Izmailov 1 Gregory Benton 1 Micah Goldblum 1 Andrew Gordon Wilson 1 1New York University. Correspondence to: Sanae Lotfi <sl8160@nyu.edu>, Andrew Gordon Wilson <andrewgw@cims.nyu.edu>.
Pseudocode No No explicit pseudocode or algorithm blocks were found in the paper.
Open Source Code No The paper does not contain an explicit statement about the release of open-source code for the described methodology or a link to a code repository.
Open Datasets Yes We investigate the correlation between the log marginal likelihood (LML) and generalization in the context of image classification using the CIFAR-10 and CIFAR-100 datasets. [...] In UCI regression tasks, we examine the performance of LML vs CLML in terms of test performance when training with limited amounts of training data. [...] we train on the Omniglot dataset and test on the EMNIST dataset.
Dataset Splits Yes We train a model on 80% of the training data, and fit the LA approximation on the same subset of the data. [...] We choose the value of T that achieves the highest BMA accuracy (average over 20 samples) on 5% of the training data. [...] The CLML is computed using a 80% 20% split of the training data as described in detail in Section D.
Hardware Specification No The paper does not provide specific hardware details (e.g., CPU/GPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions general software components like 'SGD' and 'Adam optimizer' but does not specify their versions or the versions of other key software libraries or dependencies.
Experiment Setup Yes All models were trained for 250 epochs with an SGD optimizer and an initial learning rate of 0.01. The batch-size was fixed to 128. For experiments where the prior precision was optimized, we used online optimization where the prior precision was updated every 5 epochs for 100 iterations using an Adam optimizer with an initial learning equal to 1.0.