[Paper Review] Interpretable agent communication from scratch (with a generic visual processor emerging on the side)

Roberto Dessì, Eugene Kharitonov+1 more|arXiv (Cornell University)|Jun 8, 2021

Domain Adaptation and Few-Shot Learning46 references9 citations

TL;DR

This paper proposes training two deep neural networks from scratch to develop a discrete, interpretable communication protocol for referential communication in natural images, using a contrastive learning-inspired setup. The agents learn to refer to objects—including unseen categories—via emergent symbols, while simultaneously developing high-quality generic visual features comparable to state-of-the-art self-supervised models like SimCLR.

ABSTRACT

As deep networks begin to be deployed as autonomous agents, the issue of how they can communicate with each other becomes important. Here, we train two deep nets from scratch to perform realistic referent identification through unsupervised emergent communication. We show that the largely interpretable emergent protocol allows the nets to successfully communicate even about object types they did not see at training time. The visual representations induced as a by-product of our training regime, moreover, show comparable quality, when re-used as generic visual features, to a recent self-supervised learning model. Our results provide concrete evidence of the viability of (interpretable) emergent deep net communication in a more realistic scenario than previously considered, as well as establishing an intriguing link between this field and self-supervised visual learning.

Motivation & Objective

• To develop a fully end-to-end, interpretable emergent communication protocol between deep agents without pre-trained visual features.
• To investigate whether emergent communication in referential games can yield high-quality generic visual representations.
• To evaluate the protocol's generalization to unseen object categories and its interpretability by humans.
• To bridge emergent communication and self-supervised visual representation learning through a shared training objective.
• To demonstrate that discrete, interpretable communication can emerge from scratch in realistic, large-scale image environments.

Proposed method

• The agents play a referential communication game using natural images from large-scale datasets, with no pre-trained visual encoders.
• A contrastive learning objective is used to encourage the sender to produce distinct symbols for different target objects, while the receiver must identify the correct target from a set of distractors.
• Input data augmentation is applied to improve robustness and generalization, drawing inspiration from self-supervised learning.
• The visual representations learned by the agents are extracted and evaluated as generic features on downstream classification benchmarks.
• The communication protocol is analyzed for interpretability by examining symbol-object correspondences and their consistency across categories.
• The method is implemented using the EGG framework, with training conducted end-to-end from random initialization.

Experimental results

Research questions

• RQ1 Can deep agents learn a discrete, interpretable communication protocol from scratch using only referential game performance as a training signal?
• RQ2 Do the visual representations learned during communication training generalize well to downstream vision tasks?
• RQ3 Can the emergent communication protocol generalize to object categories not seen during training?
• RQ4 How does the quality of the emergent visual features compare to those from state-of-the-art self-supervised methods?
• RQ5 To what extent is the emergent communication protocol interpretable, and can it be seen as a form of unsupervised image annotation?

Key findings

• The agents successfully learn a discrete, partially interpretable communication protocol that enables accurate referential communication, even for object categories not seen during training.
• The emergent visual representations achieve classification performance comparable to SimCLR, a self-supervised method specifically designed for visual feature learning.
• The protocol generalizes to zero-shot object categories, indicating that symbols denote semantic categories rather than just low-level image features.
• The use of input data augmentation significantly improves the robustness and quality of both communication and visual representation learning.
• The visual representations induced as a by-product of communication training are of high quality and can be reused effectively across multiple vision tasks.
• The results demonstrate a strong link between emergent communication and self-supervised visual representation learning, suggesting mutual benefits in future integration.