Drawing out of Distribution with Neuro-Symbolic Generative Models

Authors: Yichao Liang, Josh Tenenbaum, Tuan Anh Le, Siddharth N

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate Doo D on its ability to generalise across both data and tasks. We first perform zero-shot transfer from one dataset (e.g. MNIST) to another (e.g. Quickdraw), across five different datasets, and show that Doo D clearly outperforms different baselines. An analysis of the learnt representations further highlights the benefits of adopting a symbolic stroke model. We then adopt a subset of the Omniglot challenge tasks, and evaluate its ability to generate new exemplars (both unconditionally and conditionally), and perform one-shot classification, showing that Doo D matches the state of the art.
Researcher Affiliation Collaboration Yichao Liang1,2, Joshua B. Tenenbaum2, Tuan Anh Le ,3 & N. Siddharth ,4 1University of Oxford, 2MIT, 3Google, 4University of Edinburgh
Pseudocode No The paper does not contain any sections explicitly labeled "Pseudocode" or "Algorithm", nor are there structured algorithm blocks.
Open Source Code No The paper does not provide any explicit statement or link regarding the availability of open-source code for the described methodology.
Open Datasets Yes We train Doo D and Attend-Infer-Repeat (AIR) [9]2, an unsupervised part-based model, on each of five stroke-based image datasets (i) MNIST (handwritten digits) [26], (ii) EMNIST (handwritten digits and letters) [7], (iii) KMNIST (cursive Japanese characters) [6], (iv) Quickdraw (doodles) [15], and (v) Omniglot (handwritten characters from multiple alphabets) [24]
Dataset Splits No We train on MNIST and show sample reconstructions from all five datasets without fine tuning. Models are trained on each source dataset and tested on each target dataset, resulting in a 5 × 5 table for each model (Fig. 4).
Hardware Specification No The paper does not provide any specific hardware details (e.g., GPU/CPU models, processor types, memory amounts) used for running its experiments.
Software Dependencies No The paper mentions software components and libraries like Spatial Transformers, RNNs, GMM, Bernoulli, Laplace, NVIL, ADAM, and PyTorch, but does not specify their version numbers.
Experiment Setup Yes Note that we limit the maximum number of strokes to 6 throughout all experiments. estimated using the importance weighted autoencoder (IWAE) objective [3] with 200 samples (mean and standard deviation over five runs). Furthermore, in order to ensure that the ELBO objective is appropriately balanced, we employ additional weighting β for the KL-divergence over stopping criterion ot within the objective [2, 19] (see Appendix C.1 for details).