Learning Disentangled Representations with Semi-Supervised Deep Generative Models
Authors: Siddharth N, Brooks Paige, Jan-Willem van de Meent, Alban Desmaison, Noah Goodman, Pushmeet Kohli, Frank Wood, Philip Torr
NeurIPS 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our framework s ability to learn disentangled representations, both by qualitative exploration of its generative capacity, and quantitative evaluation of its discriminative ability on a variety of models and datasets. We demonstrate the efficacy of our framework on a variety of tasks, involving classification, regression, and predictive synthesis, including its ability to encode latents of variable dimensionality. |
| Researcher Affiliation | Collaboration | N. Siddharth University of Oxford nsid@robots.ox.ac.uk Brooks Paige Alan Turing Institute University of Cambridge bpaige@turing.ac.uk Jan-Willem van de Meent Northeastern University j.vandemeent@northeastern.edu Alban Desmaison University of Oxford alban@robots.ox.ac.uk Noah D. Goodman Stanford University ngoodman@stanford.edu Pushmeet Kohli Deepmind pushmeet@google.com Frank Wood University of Oxford fwood@robots.ox.ac.uk Philip H.S. Torr University of Oxford philip.torr@eng.ox.ac.uk |
| Pseudocode | No | The paper describes methods and processes but does not include a clearly labeled pseudocode block or algorithm. |
| Open Source Code | Yes | All of the above, including further details of precise parameter values and the source code, including our Py Torch-based library for specifying arbitrary graphical models in the VAE framework, is available at https://github.com/probtorch/probtorch. |
| Open Datasets | Yes | The model is tested on it s ability to classify digits and perform conditional generation on the MNIST and Google Street-View House Numbers (SVHN) datasets. Here, we use the Yale B dataset [6] as processed by Jampani et al. [12] for the results in Fig. 5. |
| Dataset Splits | No | The paper mentions sizes of supervised (M) and unsupervised (N) data points for training (e.g., “100 (out of 50000) labelled data points in the case of MNIST”), and discusses performance on a “test set”, but it does not provide explicit overall training, validation, and test dataset splits as percentages or specific counts for reproducibility across the entire dataset. |
| Hardware Specification | No | The paper states, “For all the experiments run, we choose architecture and parameters that are considered standard for the type and size of the respective datasets,” but it does not provide any specific hardware details such as GPU/CPU models, memory, or cloud instance types used for the experiments. |
| Software Dependencies | No | The paper mentions using “Py Torch” and “Ada M [16]”, but it does not provide specific version numbers for these software components or any other libraries needed to replicate the experiment. |
| Experiment Setup | Yes | For learning, we used Ada M [16] with a learning rate and momentumcorrection terms set to their default values. As for the mini batch sizes, they varied from 100-700 depending on the dataset being used and the sizes of the labelled subset Dsup. Here, scaling of the classification objective is held fixed at α = 50 (MNIST) and α = 70 (SVHN). |