Emergent Communication of Generalizations

Authors: Jesse Mu, Noah Goodman

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We examine the languages developed for our proposed communication games over two datasets: first, an artificial shape dataset which allows us to evaluate communication over cleanly defined logical concepts; second, a dataset of birds to test agents ability to learn concepts from realistic visual input.
Researcher Affiliation Academia Jesse Mu Stanford University muj@stanford.edu Noah Goodman Stanford University ngoodman@stanford.edu
Pseudocode No The paper describes the models and training procedure textually but does not include any explicit pseudocode blocks or algorithm listings.
Open Source Code Yes For full model and training details and a link to code, see Appendix A.
Open Datasets Yes We use the Shape World visual reasoning dataset [18] (Figure 2, left). We next use the Caltech-UCSD Birds dataset [40] which contains 200 classes of birds with 40 60 images (Figure 2, right).
Dataset Splits No The paper specifies test splits (e.g., '20% held out for testing', '100 classes at train and 50 at test') but does not explicitly mention a separate validation set used during training.
Hardware Specification Yes All models were trained on a single NVIDIA 2080 Ti GPU.
Software Dependencies No The paper mentions using 'PyTorch' and the 'Transformers library' but does not specify their version numbers. It also mentions the Adam optimizer [15], which is an algorithm.
Experiment Setup Yes We use the Adam optimizer [15] with learning rate 10−4, and a batch size of 64.