[Paper Review] Towards Emergent Language Symbolic Semantic Segmentation and Model Interpretability

Alberto Santamaría-Pang, James Kubricht+3 more|arXiv (Cornell University)|Jul 18, 2020

Explainable Artificial Intelligence (XAI)17 references4 citations

TL;DR

This paper proposes a symbolic semantic segmentation framework ($\text{S}^2$) that enables interpretable medical image segmentation through emergent language communication between a Sender and Receiver agent. Using a UNet-based encoder and a differentiable communication protocol, the model generates symbolic sentences describing tumor morphology, achieving performance on par with state-of-the-art methods while enabling direct interpretation of segmentation decisions.

ABSTRACT

Recent advances in methods focused on the grounding problem have resulted in techniques that can be used to construct a symbolic language associated with a specific domain. Inspired by how humans communicate complex ideas through language, we developed a generalized Symbolic Semantic ($ ext{S}^2$) framework for interpretable segmentation. Unlike adversarial models (e.g., GANs), we explicitly model cooperation between two agents, a Sender and a Receiver, that must cooperate to achieve a common goal. The Sender receives information from a high layer of a segmentation network and generates a symbolic sentence derived from a categorical distribution. The Receiver obtains the symbolic sentences and co-generates the segmentation mask. In order for the model to converge, the Sender and Receiver must learn to communicate using a private language. We apply our architecture to segment tumors in the TCGA dataset. A UNet-like architecture is used to generate input to the Sender network which produces a symbolic sentence, and a Receiver network co-generates the segmentation mask based on the sentence. Our Segmentation framework achieved similar or better performance compared with state-of-the-art segmentation methods. In addition, our results suggest direct interpretation of the symbolic sentences to discriminate between normal and tumor tissue, tumor morphology, and other image characteristics.

Motivation & Objective

• To develop a symbolic, interpretable segmentation framework that enables model transparency in medical image analysis.
• To address the interpretability gap in deep learning-based segmentation by introducing a communication-based symbolic language between agents.
• To enable direct human interpretation of model decisions through symbolic sentences describing tissue characteristics.
• To achieve competitive segmentation performance on the TCGA tumor dataset while maintaining interpretability.
• To explore emergent language as a mechanism for grounding semantic meaning in image segmentation.

Proposed method

• A UNet-like encoder network processes medical images and feeds features to a Sender network.
• The Sender generates a symbolic sentence from a categorical distribution over a learned vocabulary, representing semantic content.
• The Receiver decodes the symbolic sentence and co-generates the segmentation mask using attention and feature fusion.
• The Sender and Receiver are trained end-to-end via a differentiable communication objective, enabling emergent language learning.
• The framework uses a private, task-specific language that evolves through cooperative training to minimize segmentation loss.
• Symbolic sentences are derived from high-level features and are designed to be semantically meaningful for human interpretation.

Experimental results

Research questions

• RQ1 Can a cooperative two-agent system learn to communicate using a private symbolic language to improve segmentation interpretability?
• RQ2 Can symbolic sentences generated by the model be directly interpreted by clinicians to distinguish tumor from normal tissue?
• RQ3 Does the use of emergent language communication lead to segmentation performance comparable to state-of-the-art models?
• RQ4 To what extent can symbolic representations capture tumor morphology and other image characteristics?
• RQ5 How does the interpretability of symbolic outputs compare to black-box attention maps or feature activations?

Key findings

• The $\text{S}^2$ framework achieved segmentation performance comparable to or better than state-of-the-art methods on the TCGA tumor dataset.
• Symbolic sentences generated by the Sender network were directly interpretable by humans, enabling discrimination between normal and tumor tissue.
• The model learned to describe tumor morphology and other image characteristics through emergent language, suggesting semantic grounding.
• The cooperative training of Sender and Receiver agents enabled the emergence of a private, task-specific symbolic language.
• Interpretability was achieved without sacrificing performance, demonstrating a viable path to explainable medical AI.
• The symbolic outputs provided meaningful insights into model reasoning, unlike traditional attention maps or saliency maps.