Timezone: »

Emergent Communication: Towards Natural Language
Abhinav Gupta · Michael Noukhovitch · Cinjon Resnick · Natasha Jaques · Angelos Filos · Marie Ossenkopf · Angeliki Lazaridou · Jakob Foerster · Ryan Lowe · Douwe Kiela · Kyunghyun Cho

Sat Dec 14 08:00 AM -- 06:00 PM (PST) @ West 118 - 120
Event URL: https://sites.google.com/view/emecom2019 »

Communication is one of the most impressive human abilities but historically it has been studied in machine learning on confined datasets of natural language, and by various other fields in simple low-dimensional spaces. Recently, with the rise of deep RL methods, the questions around the emergence of communication can now be studied in new, complex multi-agent scenarios. Two previous successful workshops (2017, 2018) have gathered the community to discuss how, when, and to what end communication emerges, producing research that was later published at top ML venues such as ICLR, ICML, AAAI. Now, we wish to extend these ideas and explore a new direction: how emergent communication can become more like natural language, and what natural language understanding can learn from emergent communication.

The push towards emergent natural language is a necessary and important step in all facets of the field. For studying the evolution of human language, emerging a natural language can uncover the requirements that spurred crucial aspects of language (e.g. compositionality). When emerging communication for multi-agent scenarios, protocols may be sufficient for machine-machine interactions, but emerging a natural language is necessary for human-machine interactions. Finally, it may be possible to have truly general natural language understanding if agents learn the language through interaction as humans do. To make this progress, it is necessary to close the gap between artificial and natural language learning.

To tackle this problem, we want to take an interdisciplinary approach by inviting researchers from various fields (machine learning, game theory, evolutionary biology, linguistics, cognitive science, and programming languages) to participate and engaging them to unify the differing perspectives. We believe that the third iteration of this workshop with a novel, unexplored goal and strong commitment to diversity will allow this burgeoning field to flourish.

Sat 8:55 a.m. - 9:00 a.m. [iCal]
Introductory Remarks (Remarks)
Sat 9:00 a.m. - 9:40 a.m. [iCal]
Invited Talk - 1 (Talk)
Ted Gibson
Sat 9:45 a.m. - 10:00 a.m. [iCal]
Contributed Talk - 1 (Talk)
Mina Lee
Sat 10:00 a.m. - 10:30 a.m. [iCal]
Coffee Break / Poster Session (Poster Session)
Sat 10:30 a.m. - 11:10 a.m. [iCal]

Information-theoretic principles in semantic and pragmatic communication

Maintaining useful semantic representations of the environment and pragmatically reasoning about utterances are crucial aspects of human language. However, it is not yet clear what computational principles could give rise to human-like semantics and pragmatics in machines. In this talk, I will propose a possible answer to this open question by hypothesizing that pressure for efficient coding may underlie both abilities. First, I will argue that languages efficiently encode meanings into words by optimizing the Information Bottleneck (IB) tradeoff between the complexity and accuracy of the lexicon. This proposal is supported by cross-linguistic data from three semantic domains: names for colors, artifacts, and animals. Furthermore, it suggests that semantic systems may evolve by navigating along the IB theoretical limit via an annealing-like process. This process generates quantitative predictions, which are directly supported by an analysis of recent data documenting changes over time in the color naming system of a single language. Second, I will derive a theoretical link between optimized semantic systems and local, context-dependent interactions that involve pragmatic skills. Specifically, I will show that pressure for efficient coding may also give rise to human pragmatic reasoning, as captured by the Rational Speech Act framework. This line of work identifies information-theoretic optimization principles that characterize human semantic and pragmatic communication, and that could be used to inform artificial agents with human-like communication systems.

Noga Zaslavsky
Sat 11:15 a.m. - 11:30 a.m. [iCal]
Contributed Talk - 2 (Talk)
Alexander Cowen-Rivers
Sat 11:30 a.m. - 12:00 p.m. [iCal]
Extended Poster Session (Posters)
Travis LaCroix, Marie Ossenkopf, Mina Lee, Nicole Fitzgerald, Daniela Mihai, Jonathon Hare, Ali Zaidi, Alexander Cowen-Rivers, Alana Marzoev, Eugene Kharitonov, Luyao Yuan, Tomasz Korbak, Paul Pu Liang, Yi Ren, Roberto Dessì, Peter Potash, Shangmin Guo, Tatsunori Hashimoto, Percy Liang, Julian Zubek, Zipeng Fu, Song-Chun Zhu, Adam Lerer
Sat 2:00 p.m. - 2:40 p.m. [iCal]
Invited Talk - 3 (Talk)
Jason Eisner
Sat 2:45 p.m. - 3:00 p.m. [iCal]
Contributed Talk - 3 (Talk)
Adam Lerer
Sat 3:00 p.m. - 3:40 p.m. [iCal]
Invited Talk - 4 (Talk)
Jacob Andreas
Sat 3:45 p.m. - 4:15 p.m. [iCal]
Coffee Break / Poster Session (Poster Session)
Sat 4:15 p.m. - 4:55 p.m. [iCal]
Invited Talk - 5 (Talk)
Stefan Lee
Sat 5:00 p.m. - 5:55 p.m. [iCal]
Panel Discussion
Jacob Andreas, Ted Gibson, Stefan Lee, Noga Zaslavsky, Jason Eisner, Jürgen Schmidhuber
Sat 5:55 p.m. - 6:00 p.m. [iCal]
Closing Remarks (Remarks)

Author Information

Abhinav Gupta (Mila)
Michael Noukhovitch (Mila (Université de Montréal))

Master's student at MILA supervised by Aaron Courville and co-supervised by Yoshua Bengio

Cinjon Resnick (NYU)
Natasha Jaques (MIT)
Angelos Filos (University of Oxford)
Marie Ossenkopf (University of Kassel)
Angeliki Lazaridou (DeepMind)
Jakob Foerster (Facebook AI Research)

Jakob Foerster is a PhD student in AI at the University of Oxford under the supervision of Shimon Whiteson and Nando de Freitas. Using deep reinforcement learning he studies the emergence of communication in multi-agent AI systems. Prior to his PhD Jakob spent four years working at Google and Goldman Sachs. Previously he has also worked on a number of research projects in systems neuroscience, including work at MIT and the Weizmann Institute.

Ryan Lowe (McGill University)
Douwe Kiela (Facebook AI Research)
Kyunghyun Cho (New York University)

Kyunghyun Cho is an associate professor of computer science and data science at New York University and a research scientist at Facebook AI Research. He was a postdoctoral fellow at the Université de Montréal until summer 2015 under the supervision of Prof. Yoshua Bengio, and received PhD and MSc degrees from Aalto University early 2014 under the supervision of Prof. Juha Karhunen, Dr. Tapani Raiko and Dr. Alexander Ilin. He tries his best to find a balance among machine learning, natural language processing, and life, but almost always fails to do so.

More from the Same Authors