Shared Visual Representations of Drawing for Communication: How do different biases affect human interpretability and intent?
This work addresses the problem of improving interpretability and intent in AI-generated drawings for human-AI communication, but it is incremental, building on existing advances.
The study investigated how representational losses and inductive biases affect artificial agents' ability to produce recognizable sketches in a communication game, showing that combining pretrained encoders with biases enables effective communication and objectness emerges as a key feature.
We present an investigation into how representational losses can affect the drawings produced by artificial agents playing a communication game. Building upon recent advances, we show that a combination of powerful pretrained encoder networks, with appropriate inductive biases, can lead to agents that draw recognisable sketches, whilst still communicating well. Further, we start to develop an approach to help automatically analyse the semantic content being conveyed by a sketch and demonstrate that current approaches to inducing perceptual biases lead to a notion of objectness being a key feature despite the agent training being self-supervised.