Robust Disentanglement of a Few Factors at a Time using rPU-VAE

Authors: Benjamin Estermann, Markus Marks, Mehmet Fatih Yanik

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate how the r PU-VAE disentangles different datasets in comparison to the state-ofthe-art. We show how the performance of the r PU-VAE depends on the number of labels generated during training. We train our r PU-VAE approach with 5 random seeds and 3 leaf-runs. We note that the r PU-VAE achieves significantly higher mean disentanglement scores for all datasets used in Fig. 4.
Researcher Affiliation Academia Benjamin Estermann Institute of Neuroinformatics, ETH Zurich besterma@ethz.ch Markus Marks Institute of Neuroinformatics, ETH Zurich marksm@ethz.ch Mehmet Fatih Yanik Institute of Neuroinformatics, ETH Zurich yanik@ethz.ch
Pseudocode Yes Algorithm 1 recursive PBT-U-VAE (UDR)(r PU-VAE) 1: procedure DISENTANGLE(D) initial dataset D 2: surrogate Labels empty List() 3: labels0 PBT-U-VAE (UDR)(D) Train until convergence 4: surrogate Labels.append(labels0) 5: for i 1, Max Nr Leaf Runs do Start leaf-runs 6: leaf Labels empty List() 7: d reduce(D, labels0, i) Remove variance of the labeled factor 8: while |d| > sized min no convergence do Recursively label and reduce dataset 9: labelsi PBT-U-VAE (UDR)(d) This is one meta Epoch 10: leaf Labels.append(labelsi) 11: d reduce(d, labelsi, 0) 12: end while 13: surrogate Labels.append(leaf Labels) 14: end for 15: θ PBT-S-VAE (MIG)(D, surrogate Labels) Train final model 16: return θ 17: end procedure
Open Source Code Yes Code: https://github.com/besterma/robust_disentanglement
Open Datasets Yes For evaluation of our approaches in this study, we use the dsprites [6] and shapes3d [27] datasets, as they are some of the most commonly used in the disentanglement literature and therefore enable us to benchmark against other methods [21, 3, 23, 28, 7, 15, 12].
Dataset Splits Yes For evaluation of our approaches in this study, we use the dsprites [6] and shapes3d [27] datasets... We compute MIG as well as DCI Disentanglement as described in [7] using disentanglement-lib1.
Hardware Specification No The paper does not provide specific hardware details (exact GPU/CPU models, processor types, or memory amounts) used for running its experiments.
Software Dependencies No The paper mentions 'disentanglement-lib' but does not specify its version number, nor does it list specific version numbers for other key software components like programming languages, frameworks (e.g., PyTorch, TensorFlow), or other libraries.
Experiment Setup No The paper discusses hyperparameter sensitivity and tuning generally, and refers to 'Supplementary section B' for how hyperparameters developed, but it does not provide specific hyperparameter values or detailed training configurations (e.g., learning rate, batch size, number of epochs) within the main text provided.