Diversity vs. Recognizability: Human-like generalization in one-shot generative models
Authors: Victor Boutin, Lakshya Singhal, Xavier Thomas, Thomas Serre
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here, we propose a new framework to evaluate one-shot generative models along two axes: sample recognizability vs. diversity (i.e., intra-class variability). Using this framework, we perform a systematic evaluation of representative one-shot generative models on the Omniglot handwritten dataset. We first show that GAN-like and VAE-like models fall on opposite ends of the diversity-recognizability space. Extensive analyses of the effect of key model parameters further revealed that spatial attention and context integration have a linear contribution to the diversity-recognizability trade-off. |
| Researcher Affiliation | Academia | 1 Artificial and Natural Intelligence Toulouse Institute, Université de Toulouse, France 2 Carney Institute for Brain Science, Dpt. of Cognitive Linguistic & Psychological Sciences Brown University, Providence, RI 02912 |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Our code could be found at https://github.com/serre-lab/diversity_vs_recognizability. |
| Open Datasets | Yes | In this article, we use the Omniglot dataset [32] with a weak generalization split [42]. |
| Dataset Splits | No | The paper describes training and testing splits, but it does not explicitly provide details about a separate validation dataset split (e.g., percentages, sample counts, or how it was created). |
| Hardware Specification | No | The paper mentions that hardware specifications are in the Supplementary Information at section S18, but the main paper does not contain specific details like GPU models or CPU types. |
| Software Dependencies | No | The paper mentions various models and networks (e.g., VAE, GAN, Inception v3 Net), but it does not provide specific version numbers for any software libraries or dependencies used (e.g., PyTorch, TensorFlow, Python). |
| Experiment Setup | Yes | For all algorithms listed in section 2.2 we have explored different hyper-parameters (see section 4.2 for more details). ... We evaluate the effect of the context on the position of the VAE-NS models on the diversity-recognizability space by varying the number of samples used to compute the context statistics during the training phase (from 2 to 20 samples). ... We have varied the number of attentional steps from 20 to 90. ... One can operate such a modulation by changing the β coefficient in the ELBO loss function [24]. |