Tomographic Auto-Encoder: Unsupervised Bayesian Recovery of Corrupted Data

Authors: Francesco Tonolini, Pablo Garcia Moreno, Andreas Damianou, Roderick Murray-Smith

ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We test our model in a data recovery task under the common setting of missing values and noise, demonstrating superior performance to existing variational methods for imputation and de-noising with different real data sets. We further show higher classification accuracy after imputation, proving the advantage of propagating uncertainty to downstream tasks with our model.
Researcher Affiliation Collaboration Francesco Tonolini School of Computing Science University of Glasgow Glasgow, UK 2402432t@student.gla.ac.uk Pablo Garcia Moreno Amazon London, UK morepabl@amazon.co.uk Andreas Damianou Amazon London, UK damianou@amazon.co.uk Roderick Murray-Smith School of Computing Science University of Glasgow Glasgow, UK roderick.murray-smith@glasgow.ac.uk
Pseudocode Yes Algorithm 1 Training the TAE Model
Open Source Code No The paper does not contain any explicit statement about releasing source code or a link to a code repository for the described methodology.
Open Datasets Yes We corrupt the MNIST dataset (Deng, 2012), We further evaluate our TAE with Fashion-MNIST 28 28 grey-scale images of clothing (Xiao et al., 2017), and the UCI HAR dataset, which consists of filtered accelerometer signals from mobile phones worn by different people during common activities (Anguita et al., 2012).
Dataset Splits No The paper mentions repeating experiments and using subsets of data for classification tasks (e.g., '10,000 examples... 1,000 of these are labelled... classify the remaining 9,000'), but does not provide explicit train/validation/test dataset splits with percentages or sample counts for the main model training.
Hardware Specification Yes All experiments were performed using a Titan X GPU.
Software Dependencies No The paper mentions 'Tensorflow' but does not specify a version number or other software dependencies with their versions.
Experiment Setup Yes Common parameters are as follows: 500, 000 iterations with the ADAM optimiser in Tensorflow, an initial training of 2e-4 and batch size of 20. The hyper-parameters specific to the TAE are instead: γ initially set to 0.01 and then linearly increased to 1 between 50, 000 and 100, 000 iterations, λ = 2 and C = 10.