Development of the agentive self: Critical components in the emerging ability of action prediction and goal anticipation

以与目标相关的方式控制自己的行动，以及理解他人行动背后的目标和意图的能力是能动自我的重要方面。这两个方面都在婴儿期发展，并依赖于迄今为止获得的自己的代理经验。有人假设，自己的行动的认知表征部分用于计划自己的行动和理解他人的行动。然而，当观察到的运动映射到认知动作表示，具有挑战性的问题的对应关系，角度和电机推理需要解决。尽管镜像神经元系统（MNS）被认为是一个关键的角色，但实际的编码和计算过程以及它们的个体发育仍然是难以捉摸的。计划中的项目旨在通过将发展心理学的见解和进一步的实验评估与面向机器学习的认知建模相结合来填补这一解释性空白。这种跨学科的合作有望以双向的方式带来好处：发展心理学将通过行动理解的认知发展的功能性计算模型来增强。机器学习和认知系统研究将受益于识别促进行动理解出现的归纳偏见。在对婴儿进行的眼动追踪和脑电图研究中，并在建模工作的推动下，该项目将进一步详细评估哪些提示和提示组合（例如人类视觉外观，自我推进，显著动作效果的产生，自己的动作体验）与婴儿预测观察到的动作目标的能力最相关。计算模型将把我们目前的生物运动模型与我们的事件预测认知理论结合起来。例如，目前的模型预计，感知突出的最终目标将支持预期的行动观察。通过对具体场景进行建模，我们还将生成更具体的行为预测。总的来说，我们希望回答关键的发展和认知科学问题。例如，对于哪些类型的观察到的动作，眼动跟踪和脑电图衍生的MNS活动信号是可检测的？不同年龄段婴儿的自身行为经验或对他人行为的观察是否会影响其对后续行为的理解？代理线索增强能促进行动理解的计算模型的学习吗？因此，该项目将有助于SPP 2134的主题重点，将（神经）认知建模产生的预测与发展心理学的见解交织在一起，以促进对婴儿如何（i）计划和控制自己的目标相关行动以及（ii）预测行动目标并推断他人的潜在意图的理解。总的来说，该项目将进一步阐明代理自我，MNS的发展，以及由此产生的社会能力。

The abilities to control one’s own actions in a goal-related way and to understand the goals and intentions behind the actions of others are important aspects of the agentive self. Both aspects develop during infancy and depend on the so-far acquired own agentive experience. It has been hypothesized that cognitive representations of own actions are partially used to plan own actions and to understand the actions of others. However, when mapping observed motions onto cognitive action representations, challenging problems of correspondence, perspective, and motor inference need to be solved. Although a critical role of the mirror neuron system (MNS) is supposed, the actual encodings and computational processes involved as well as their ontogenetic development remain elusive. The planned project seeks to fill this explanatory gap by combining insights and further experimental evaluations from developmental psychology with machine learning-oriented cognitive modeling. This interdisciplinary collaboration promises benefits in a bidirectional manner: Developmental psychology will be augmented with a functional, computational model of the cognitive development of action understanding. Machine-learning and cognitive-systems research will profit from the identification of inductive biases that foster the emergence of action understanding. In eye-tracking and EEG studies with infants and driven by the modeling efforts, the project will assess in further detail which cues and cue combinations of agency (e.g. human visual appearance, self-propelledness, production of salient action-effects, own action experience) are most relevant for infants’ ability to anticipate the goals of observed actions. The computational models will combine our current biological-motion model with our theory of event-predictive cognition. The current model expects, for example, that perceptual highlighting the final goal will support anticipatory action observations. By modeling the concrete scenarios, we will also generate more concrete behavioral predictions. Overall, we expect to answer critical developmental and cognitive-science questions. For example, for which types of observed actions will eye-tracking- and EEG-derived signals for MNS activity be detectable? Do own action experiences or observations of others’ actions influence subsequent action understanding in infants of different ages? Can agency-cue augmentations facilitate the learning of computational models of action understanding? Thus, the project will contribute to the thematic focus of the SPP 2134 by interweaving predictions generated by (neuro-)cognitive modeling with insights from developmental psychology to foster understanding on how infants (i) plan and control own goal-related actions as well as (ii) anticipate action goals and infer the underlying intentions of others. Overall, the project will shed further light on the development of the agentive self, the MNS, and the resulting social competencies.