It's About Time: Temporal References in Emergent Communication
Olaf Lipinski, Adam J. Sobey, Federico Cerutti, Timothy J. Norman
TL;DR
This paper investigates how agents develop temporal references in emergent communication, finding that architectural changes, specifically batching methods, are crucial for their emergence, which enhances communication efficiency.
Contribution
It demonstrates that minimal architectural modifications enable temporal references to emerge, highlighting the importance of design choices over loss function adjustments.
Findings
Over 95% of agents with modified batching develop temporal references
Architectural changes are necessary for temporal reference emergence
Temporal referencing improves communication efficiency and coding optimality
Abstract
Emergent communication studies the development of language between autonomous agents, aiming to improve understanding of natural language evolution and increase communication efficiency. While temporal aspects of language have been considered in computational linguistics, there has been no research on temporal references in emergent communication. This paper addresses this gap, by exploring how agents communicate about temporal relationships. We analyse three potential influences for the emergence of temporal references: environmental, external, and architectural changes. Our experiments demonstrate that altering the loss function is insufficient for temporal references to emerge; rather, architectural changes are necessary. However, a minimal change in agent architecture, using a different batching method, allows the emergence of temporal references. This modified design is compared…
Peer Reviews
Decision·Submitted to ICLR 2024
1. This paper introduces temporal reference into the emergent communication field, which is an important topic worth exploring. It leverages the sequential position of objects within the batch as a means of encoding temporal information, thereby modifying traditional referential games and their associated training frameworks. 2. The paper demonstrates a comprehensive design of the experiments. It encompasses various modifications to modeling architectures, incorporates multiple environments, an
1. Example demonstration: the authors present multiple examples using integer arrays in the game design (section 2.3), agent architecture (figure 1), how to compute the temporality metric (section 3.1), as well as the sent messages (section 3.3), which I believe are of different meanings. However, using integer arrays in all the examples can make the demonstration more confusing. Different symbols used in different cases would be better. For example, objects as names, and messages using alphabet
- (major) Temporal reference is a pertinent feature of language, and therefore, it is good to use an environment with multiple timesteps with additional temporal structure (i.e., recently seen object are more likely to appear again). - (minor) The Temporal Referential Game (TRG) is an appropriate extension to the regular referential game. It minimally adds a temporal aspect to the game without changing it too much, making it good for pioneering basic concepts about temporal referenc
This paper does not make a substantial contribution to the field. While the TRG could elicit potentially interesting behaviors, the primary thrust of this paper is that it introduces a new architecture which is able to solve this game without much further analysis about temporality. The problem here is that the new architecture is, to the best of my knowledge, just an LSTM where each timestep is a single round of the referential game. This, to me, seems like an appropriate baseline with which
- It is the first work to investigate the temporal references in emergent communication, during which it defines an environment and introduces an architecture of the agent. The subject matter is interesting and vital. - Despite the limitations listed below, the experiments to a certain extent provide a response to the major research question of this study and support for the conclusions.
1. The notations in this paper need to be further standardised and unified. The current ones sometimes make readers (at least me) hard to follow. Here are the ones that I have spotted: (1) a vector in this paper sometimes means the feature vector of a target and sometimes a target sequence. Both of them can have arbitrary lengths and often appear without further explanation. (2) things like x \in V are not conditions in equation 2, IMO, they should not connected with the "real" conditions with A
Code & Models
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsLanguage and cultural evolution · Speech and dialogue systems