On the role of population heterogeneity in emergent communication
Authors: Mathieu Rita, Florian Strub, Jean-Bastien Grill, Olivier Pietquin, Emmanuel Dupoux
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper, we explore emergent language properties by varying agent population size in the speaker-listener Lewis Game. After reproducing the experimental difference, we challenge the simulation assumption that the agent community is homogeneous. We then investigate how speaker-listener asymmetry alters language structure through the analysis a potential diversity factor: learning speed. From then, we leverage this observation to control population heterogeneity without introducing confounding factors. We finally show that introducing such training speed heterogeneities naturally sort out the initial contradiction: larger simulated communities start developing more stable and structured languages. |
| Researcher Affiliation | Collaboration | Mathieu Rita INRIA, Paris mathieu.rita@inria.fr Florian Strub Deep Mind fstrub@deepmind.com Jean-Bastien Grill Deep Mind jbgrill@deepmind.com Olivier Pietquin Google Research, Brain Team pietquin@google.com Emmanuel Dupoux EHESS,ENS-PSL,CNRS,INRIA Meta AI Research emmanuel.dupoux@gmail.com |
| Pseudocode | No | The paper does not contain any section or figure explicitly labeled 'Pseudocode' or 'Algorithm', nor any structured, code-like blocks describing a procedure. |
| Open Source Code | Yes | The code is available at https://github.com/Mathieu Rita/Population. |
| Open Datasets | No | For each new run, we generate objects in VK uniformly at random, and we split them into a train and a test sets, which respectively contain 80% and 20% of the objects. No specific public dataset is named or linked, nor is the generated dataset made available. |
| Dataset Splits | No | For each run, we generate objects in VK uniformly at random, and we split them into a train and a test sets, which respectively contain 80% and 20% of the objects. A separate validation split is not explicitly mentioned. |
| Hardware Specification | No | The paper mentions gaining 'access to the HPC resources of IDRIS under the allocation 2021-AD011012278 made by GENCI' but does not provide specific hardware details like GPU models, CPU models, or memory specifications. |
| Software Dependencies | Yes | Sci Py 1.0: Fundamental Algorithms for Scientific Computing in Python. Nature Methods, 17:261 272, 2020. doi: 10.1038/s41592-019-0686-2. |
| Experiment Setup | Yes | In the main paper, we use |K| = 4 attributes with |V| = 4 values. ...Finally, we use a vocabulary size of |W| = 20 and a message length T = 10. ... hidden size of 128. Optimization: For both agents, we use a Adam optimizer with a learning rate of 5e10 3, β1 = 0.9 and β2 = 0.999 and a training batch size of 1024 when optimizing their respective loss. For the speaker, we set the entropy coefficient of 0.02. |