Postdoctoral Fellowship: SPRF: A Comprehensive Modeling Framework for Semantic Memory Search

该奖项是美国国家科学基金会社会、行为和经济科学（SBE）博士后研究奖学金（SPRF）计划的一部分。SPRF计划的目标是为学术界、工业界或私营部门和政府的科学事业准备有前途的早期职业博士级科学家。SPRF奖励包括在知名科学家的赞助下进行为期两年的培训，并鼓励博士后进行独立研究。美国国家科学基金会寻求促进科学界各阶层的科学家，包括那些未被充分代表的群体的科学家，参与其研究项目和活动；博士后阶段被认为是实现这一目标的一个重要的专业发展阶段。每个博士后必须解决各自学科领域的重要科学问题。在宾夕法尼亚大学苏迪普·巴蒂亚博士的赞助下，这个博士后奖学金奖支持一位早期职业科学家研究成年人如何搜索他们的一般知识来回答简单的问题。人们每天在交谈、产生新想法、认识熟悉的和新奇的物体等时，都会搜索自己的常识。拟议研究的目的是建立一个计算机模型，以与人类相同的方式搜索知识库。开发这样一个计算机模型，可以正式说明大脑在完成某些行为时所进行的过程，这些模型可以作为人类思维的定量数学理论。这个项目提出了一种新颖而统一的计算方法来模拟人们如何从他们所知道的东西中产生想法。我们的目标是通过收集人类对这些简单任务的反应来评估这种方法，并检查我们模型的基本变体如何很好地预测这些反应。一旦我们有了一个很好的想法，我们的模型与人类搜索记忆的方式是一致的，我们将能够在更多的临床人群中检查记忆。目前的提案详细介绍了一个项目，旨在建立一个语义知识检索的计算模型，并使用自然的人体实验来测试模型假设。提出的模型旨在阐明人们如何从记忆中检索知识，这是一个受到认知科学家和心理学家相当关注的重要问题。然而，研究人员尚未开发出语义记忆搜索的认知过程模型，以参数化知识检索的机制，并预测人类参与者对任意开放式知识检索提示所列出的概念、特征或关系的序列。例如，有数百万个潜在的特征和关系可以描述一个给定的目标概念。明确这些特征和关系，并对基于这些特征和关系的记忆搜索过程进行建模，对研究人员提出了重大的理论和技术挑战。我们计划使用被称为变压器网络的人工智能新技术来应对这些挑战。我们将使用现有的特征规范数据集来训练网络，以预测数百万个不同的特征和关系中哪些是共同概念。然后，我们将使用这些训练过的网络来生成一个知识库，该知识库由我们提出的语义记忆搜索模型的表示构成。随后，我们将在广泛的开放式知识检索任务上使用个人层面的参数模型拟合来评估我们的模型，包括语义流畅性、特征生成和模拟生成。如果成功，该项目将提供一个新的理论范式，将语义认知的计算模型与记忆搜索的认知过程模型相结合。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This award was provided as part of NSF's Social, Behavioral and Economic Sciences (SBE) Postdoctoral Research Fellowships (SPRF) program. The goal of the SPRF program is to prepare promising, early career doctoral-level scientists for scientific careers in academia, industry or private sector, and government. SPRF awards involve two years of training under the sponsorship of established scientists and encourage Postdoctoral Fellows to perform independent research. NSF seeks to promote the participation of scientists from all segments of the scientific community, including those from underrepresented groups, in its research programs and activities; the postdoctoral period is considered to be an important level of professional development in attaining this goal. Each Postdoctoral Fellow must address important scientific questions that advance their respective disciplinary fields. Under the sponsorship of Dr. Sudeep Bhatia at the University of Pennsylvania, this postdoctoral fellowship award supports an early career scientist investigating how adults search their general knowledge to answer simple questions. People search their general knowledge every day when they converse, come up with new ideas, recognize familiar and novel objects, etc. The aim of the proposed research is to build a computer model that searchers a store of knowledge in the same way that humans do. Developing such a computer model enables formal specification of the processes that the mind carries out when it accomplishes certain behaviors, and these models serve as quantitative, mathematical theories of human thought. This project presents a novel and unified computational approach for modeling how people generate ideas from what they know. We aim to evaluate this approach by collecting human responses to these simple tasks and examining how well basic variants of our model can predict these responses. Once we have a good idea that our model coheres with the way that humans search their memory, we will be able examine the memory in more clinical populations. The current proposal details a project which aims build a computational model of semantic knowledge retrieval and test model assumptions using naturalistic human experiments. The proposed model aims to clarify how people retrieve knowledge from memory, an important issue that has received considerable attention from cognitive scientists and psychologists. However, researchers have not yet developed cognitive process models of semantic memory search that can parameterize mechanisms involved in knowledge retrieval, and predict sequences of concepts, features, or relations listed by human participants, in response to arbitrary open-ended knowledge retrieval prompts. There are, for example, millions of potential features and relations that could describe a given target concept. Specifying these features and relations, and modeling the memory search processes that operate on these features and relations, poses significant theoretical and technical challenges for researchers. We plan to address these challenges using new techniques in artificial intelligence known as transformer networks. We will use existing feature norm datasets to train the networks to predict which of millions of distinct features and relations hold for common concepts. We will then use these trained networks to generate a knowledge base constituting the representations over which our proposed models of semantic memory search operate. We will subsequently evaluate our models using individual-level parametric model fitting on a wide range of open-ended knowledge retrieval tasks, including semantic fluency, feature generation, and analog generation. If successful, this project will offer a novel theoretical paradigm that integrates computational models of semantic cognition with cognitive process models of memory search.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.