Adaptive Latent Feature Sharing for Piecewise Linear Dimensionality Reduction
Authors: Adam Farooq, Yordan P. Raykov, Petar Raykov, Max A. Little
JMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | This section provides some empirical results on the performance of the proposed variants of PCA and FA applied to data visualization, data whitening and blind source separation. The methods are evaluated on different kinds of synthetic data, images of handwritten digits from MNIST, images of objects from the Coil-20 data set, and functional magnetic resonance imaging (f MRI) data. ... The utility of the proposed a PPCA and a FA models is demonstrated on standard tasks such as feature learning, data visualization, and data whitening. |
| Researcher Affiliation | Academia | Adam Farooq EMAIL Department of Mathematics Aston University Birmingham, UK; Yordan P. Raykov EMAIL School of Mathematical Sciences University of Nottingham Nottingham, UK; Petar Raykov EMAIL MRC Cognition and Brain Sciences Unit University of Cambridge Cambridge, UK; Max A. Little EMAIL Department of Computer Science University of Birmingham Birmingham, UK |
| Pseudocode | Yes | Algorithm 1 EM algorithm for parametric adaptive factor (a FA) analysis. ... Algorithm 3 Inference in parametric a PPCA using Gibbs sampling. |
| Open Source Code | No | The paper describes the use of third-party tools like the 'Pymanopt toolbox (Townsend et al., 2016)' for specific parts of their implementation, and mentions 'STAN (Carpenter et al., 2017)' as a potential MCMC platform. However, it does not contain any explicit statement about the release of source code for the methodology described in the paper, nor does it provide a direct link to a code repository for their own implementation of aFA or aPPCA. |
| Open Datasets | Yes | The methods are evaluated on different kinds of synthetic data, images of handwritten digits from MNIST, images of objects from the Coil-20 data set, and functional magnetic resonance imaging (f MRI) data. ... dataset from the Columbia University Image Library (COIL-20) (Nene et al., 1996). ... f MRI data collected from a single participant while exposed to continuous visual stimuli (Raykov et al., 2021). |
| Dataset Splits | Yes | Parameters for each model were estimated using 80% of the data set, and the resulting model was tested on the remaining 20%; average of the mean square out-of-sample reconstruction error over 20 different experiments, is reported. ... evaluated using 10-fold cross-validation on 10, 000-image subset of the 784 D MNIST data set. |
| Hardware Specification | No | The paper describes the equipment used for fMRI data acquisition ('Using a 3T Siemens scanner, a whole brain image with a voxel resolution of 2 2 2 mm was acquired every 0.8 seconds.'), but it does not provide specific details about the computational hardware (e.g., CPU, GPU models, memory, or cloud instances) used for running the experiments or training the models. |
| Software Dependencies | No | The paper mentions the use of 'Pymanopt toolbox (Townsend et al., 2016)' and that 'STAN (Carpenter et al., 2017)' could be used, but it does not provide specific version numbers for these or any other software components, which is required for a reproducible description of ancillary software. |
| Experiment Setup | Yes | The total number of unique factors is set to K = 100 and the number L of observation-specific factors is set to maximize the factor profiles of the different digits. ... For a PPCA, we have kept L = K 1 for simplicity. ... For a PPCA we used K = 500 unique components and constraint of L = 200 components, which were selected to achieve component similarity with the benchmark and enable visually intuitive comparisons. ... We train a simple multilayer perceptron with one hidden layer with a softmax activation function evaluated using 10-fold cross-validation on 10, 000-image subset of the 784 D MNIST data set. |