Learning Knowledge of Agency

Distinguishing Self, Other, and Autonomy From Visual Feedback: A Combined Correlation and Acceleration Transfer Analysis

“…From early development, self and other distinctions are fundamental to the focus the learning (reduce state space dimensionality) of forward model acquisition. From a computational perspective, discriminating between self and other features from visual feedback is often addressed through models (Brody et al, 2017 ; Sánchez-Fibla et al, 2017b ; Thomas et al, 2017 ; Pertsch et al, 2018 ; Rybkin et al, 2018 ), which either require a multitude of parameters (deep learning approaches like Rabinowitz et al, 2018 ) or rely on the predictive coding hypothesis, requiring a forward model to be able to check the matching between current and predicted states (Fairhurst et al, 2019 ). In this study, we have approached this problem from a principled perspective, identifying minimum requirements to solve the problem of deciphering which features of the visual scene correspond to the self and which of them correspond to other entities in the scene, via a correlation analysis of velocity signals, that we have found to be sufficient.…”

“…Acquiring ToM abilities requires labeling and clustering the SM data stream of the interaction of the agent in terms of which visual features belong to its own body, which ones belong to other entities, which ones can be controlled by its actuators, and which ones can be controlled by themselves or are passive and need others to move. The problem of deciphering self from others in robotics and AI has been addressed by several computational models in studies such as Brody et al ( 2017 ), Thomas et al ( 2017 ), Sánchez-Fibla et al ( 2017b ), Rybkin et al ( 2018 ), and Pertsch et al ( 2018 ). We approach the labeling problem from the perspective of identifying what are the minimal requirements to distinguish self, other, and autonomous or passive entities from visual feedback alone.…”