In this proposal, the candidate introduces the idea to perform his work in a hot research topic in the field of Machine Learning and Natural language processing, such as Emergent communication. The broad aim of the project is to let artificial agents interact with each other and develop a language that shows similar syntactic measures to the human one, e.g., compositionality.
Identifying a solid learning framework is the first critical step to give agents the capability of building up their own communication. Therefore, we select the EGG environment which has been used in the literature as our primary environment and extend it. The latter implements the referential game setting, first introduced by David Lewis, which is considered a stable environment to study language emergent for a wide variety of agents.
The second critical step is finding a set of evaluation metrics to understand what kind of language is being developed by the artificial agents. We leverage recent works that focus their attention on implementing such metrics and analyze our results accordingly.
In our proposed work, we adopt the Referential game and tune the environment complexity to study different visual settings¿ effect on the emerging language. We base our environment on the popular EGG framework [12] developed at Facebook on which multiple papers are found[1;2;3;10].
We will augment the sender architecture by integrating knowledge of the context from which the target is sampled. This additional information closes the gap between the artificial environment and what humans experienced when first developing communication, thus emerging a more human-like language in the syntax dimension.
[1] Anti-efficient encoding in emergent communication. Rahma Chaabouni, Eugene Kharitonov, Emmanuel Dupoux, Marco Baroni. NeurIPS 2019.
[2] Focus on What¿s Informative and Ignore What¿s not: Communication Strategies in a Referential Game. Roberto Dessì, Diane Bouchacourt, Davide Crepaldi, Marco Baroni. NeurIPS Workshop on Emergent Communication 2019. A Sender/Receiver game where the Sender sees a target as a vector of discrete properties (e.g. [2, 4, 3, 1] for a game with four dimensions) and the Receiver has to recognize the target among a lineup of target+distractor(s).
[3] Entropy Minimization In Emergent Languages. Eugene Kharitonov, Rahma Chaabouni, Diane Bouchacourt, Marco Baroni. ICML 2020. egg/zoo/language_bottleneck contains a set of games that study the information bottleneck property of the discrete communication channel. This property is illustrated in an EGG-based example of MNIST-based style transfer without an adversary
[10] R. Chaabouni, E. Kharitonov, D. Bouchacourt, E. Dupoux, and M. Baroni.Compositionality and generalization in emergent languages.arXiv preprintarXiv:2004.09124, 2020.
[12] Kharitonov, Eugene, et al. "EGG: a toolkit for research on Emergence of lanGuage in Games." arXiv preprint arXiv:1907.00852 (2019).