Towards Conceptual Compression
Authors: Karol Gregor, Frederic Besse, Danilo Jimenez Rezende, Ivo Danihelka, Daan Wierstra
NeurIPS 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We introduce convolutional DRAW, a homogeneous deep generative model achieving state-of-the-art performance in latent variable image modeling. The subsequent sections then describe the algorithm and present results both on generation quality and compression. |
| Researcher Affiliation | Industry | Karol Gregor Google Deep Mind karolg@google.com Frederic Besse Google Deep Mind fbesse@google.com Danilo Jimenez Rezende Google Deep Mind danilor@google.com Ivo Danihelka Google Deep Mind danihelka@google.com Daan Wierstra Google Deep Mind wierstra@google.com |
| Pseudocode | No | The paper presents the algorithm's equations (1-12) and explains them (e.g., 'Below we present the equations for a one layer system... We follow the computations and explain them and the variables as we go along.'), but it does not include a distinct block labeled 'Pseudocode' or 'Algorithm', nor are the steps formatted like structured code. |
| Open Source Code | No | The paper does not include any explicit statements about releasing the source code for the methodology described, nor does it provide a link to a code repository. |
| Open Datasets | Yes | We trained the models on Cifar-10, Omniglot and Image Net with 320, 160 and 160 LSTM feature maps, respectively. We use the version of Image Net presented in (van den Oord et al., 2016). Omniglot The recently introduced Omniglot dataset Lake et al. (2015) |
| Dataset Splits | No | The paper mentions 'Test set performance' in Table 1 for Omniglot, CIFAR-10, and Image Net, implying the use of test splits. However, it does not explicitly provide specific percentages, sample counts, or detailed methodologies for training, validation, and test splits needed for reproduction (e.g., '80/10/10 split' or '5-fold cross-validation'). It implicitly relies on standard splits for these datasets. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU or CPU models, memory amounts, or cloud instance types used for running its experiments. It only mentions 'LSTM feature maps' which are model parameters. |
| Software Dependencies | No | The paper mentions using 'Adam optimization (Kingma & Ba, 2014)' and 'Long Short-Term Memory networks (LSTM; Hochreiter & Schmidhuber, 1997)', but it does not specify any version numbers for these or other software dependencies like programming languages, libraries, or frameworks (e.g., Python version, TensorFlow/PyTorch version). |
| Experiment Setup | Yes | All models (except otherwise speciļ¬ed) were single-layer, with the number of DRAW time steps nt = 32, a kernel size of 5 5, and stride 2 convolutions between input layers and hidden layers with 12 latent feature maps. We trained the models on Cifar-10, Omniglot and Image Net with 320, 160 and 160 LSTM feature maps, respectively. We train the network with Adam optimization (Kingma & Ba, 2014) with learning rate 5 10 4. |