A Systematic Exploration of when Elements of Episodic Memory Begin to Emerge in Young Children

几十年来，人们一直在研究情景记忆的潜在机制，包括它的出现和个体发生。许多关于情景记忆出现的早期研究声称，它直到4岁左右才出现（例如，Tulving, 2005）。然而，最近的研究，对更年幼的儿童使用适应性EM测量，发现情节记忆在3岁时就存在（例如，Bauer等人，2013年），甚至在婴儿2岁时就存在（Bauer和Leventon， 2013年）。对这些发现的表面解释仅仅意味着先前的研究表明了EM出现的不确定性。然而，这些发现不应该被并行考虑，因为它们采用了不同的任务（例如，图片序列记忆任务和诱导模仿），虽然这些已知是EM的测量方法，但我们必须小心假设它们测量的是相同的心理过程。此外，情景记忆可能要到4岁才完全形成，但这可能是由于构成情景记忆的认知能力的逐渐出现和发展造成的。这些发现只是延续了情景记忆不应该被视为一个单一的结构，而是一个多方面的过程的概念。以系统的方式实施和使用情景记忆的多重评估可以让我们更好地了解这些不同的过程在早期什么时候开始出现，以及它们发展的轨迹。这种用于评估EM元素的“分而攻之”设计（crystal & Suddendorf, 2019）在与语言前或早期参与者合作时尤为重要，因为试图创建一个单一的、统一的整体EM衡量标准可能需要复杂的认知水平和语言能力，而这些水平在这个阶段尚未完全发展。如果这种电池可以有效地为非常年幼的儿童开发，那么这些评估就可以用于测试人工智能代理。最近的一项研究，即动物-人工智能试验台（Crosby等人，2020），采用了一种比较认知设计，将人工智能代理放置在3D虚拟环境中，在那里它们可以接受一系列认知任务的测试，这些任务之前已经由动物成功完成。然而，值得注意的是，对人工智能代理中EM所需的认知过程的评估仍有待开发，这可以为将人工智能集成到社会和专业领域的快速发展工作提供宝贵的见解。现有文献的简要概述揭示了我在这个项目中要克服的两个主要空白：婴儿期/幼儿期情景记忆的系统个体发生，以及开发用于测试人工智能代理情景记忆元素的稳健有效评估。因此，研究目标如下：1)实施现有的并开发新的EM测试，最终为非常年幼的儿童进行一系列评估。2)使用任务表现数据为学龄前婴儿和人工智能代理开发额外的EM测试。这项研究将有助于全面了解认知如何发展到最终形成完整的、重建的情景记忆。幼儿情景样认知进化的时间表将为EM出现的知识库增加新的见解。开发一种测试人工智能情景记忆的媒介，也将使我们能够在这一指标上进行第一次直接的人类与人工智能的比较。

There have been decades of research into the underlying mechanisms of episodic memory including its' emergence and ontogeny. Much of the early research into the emergence of episodic memory claims that it does not appear until around the age of 4 (e.g., Tulving, 2005). However, more recent research, using adapted measures of EM for even younger children, have found evidence for episodic memory present at 3 years (e.g., Bauer et al, 2013) and even by the 2nd year of infancy (Bauer & Leventon, 2013).A surface-level interpretation of these findings would merely imply that previous research illustrates uncertainties regarding the emergence of EM. However, these findings should not be considered in parallel with one another as they adopt the use of different tasks (e.g., picture sequence memory task and elicited imitation) and whilst these are known to be measures of EM, we must be careful to assume that they are measuring identical psychological processes. Furthermore, it may be the case that episodic memory does not fully emerge until the age of 4, however this is likely due to the gradual emergence and development of constituent cognitive abilities that contribute to complete episodic memory. Such discoveries only perpetuate the notion that episodic memory should be approached not as a unitary construct, but rather a multifaceted process. Implementing and using multiple assessments of episodic memory in a systematic manner can allow us to gain a greater understanding of when these different processes begin to emerge in earlier years and the trajectory at which they develop. This 'divide-and-conquer' design for assessing elements of EM (Cystal & Suddendorf, 2019) is particularly essential when working with pre- or early-verbal participants, as attempting to create a single, unified measure of wholistic EM may require complex levels of cognition and language abilities that have not yet been fully developed at this stage.If such a battery can be effectively developed for very young children, these assessments could then be adapted to test AI agents. A recent line of research, the Animal-AI testbed (Crosby et al., 2020), adopts a comparative cognition design, placing an AI agent in a 3D virtual environment where they can be tested on a range of cognitive tasks that have been previously successfully completed by animals. Notably, however, there remains to have been developed an assessment of the cognitive processes required for EM in AI agents, which could provide invaluable insight into the rapidly evolving work surrounding the integration of Artificial Intelligence into social and professional spheres.This brief outline of existing literature reveals two overarching lacunae that I aim to overcome in this project: a systematic ontogeny of episodic memory in infancy/early childhood, and the development of robust and valid assessments for testing elements of episodic memory in AI agents. As such, the research aims are as follows: 1) To implement existing, and develop novel, tests of EM culminating in a battery of assessments for very young children 2) To use the data on task performance to develop additional tests of EM for preverbal infants and AI agents This research will contribute to a comprehensive timeline of how cognition develops to culminate in full, reconstructive, episodic memory. A timeline of this evolvement of episodic-like cognition in young children would add novel insight to the knowledge base of EM emergence. The development of a medium through which to test episodic memory in artificial intelligence would also allow us to conduct the first direct human-AI comparison on this measure.