REU Site: Beyond Language: Training to Create and Share Vector Embeddings across Applications

这个位于北德克萨斯大学的为期三年的本科生研究体验(REU)网站将在每年夏天为10名学生提供10周的支持，并培训他们构建能够跨域交流所获得知识的现代人工智能(AI)系统。最先进的人工智能系统用于具有挑战性的目标，如视觉对象识别、语音识别和理解自然语言，需要进行广泛的培训，将许多学习到的知识锁定在所构建的人工智能系统的结构中。然而，本课程的学生将专注于创建和利用训练有素的人工智能系统来重新表示信息，以便保留这些系统所学到的细微差别理解，并随时可供其他应用程序访问。该REU汇集了一个跨学科团队，以支持展示人工智能系统的好处的项目，这些系统旨在直接提供和接收推断知识。在课程开始时，每个学生将从所提供的许多选项中确定要使用的应用程序域和外部顾问。学生还将参加一个长期的、独特的人工智能暑期研究项目，整合当前的大学生和外部REU学生，以促进跨部门和学生专业的合作。具体地说，该REU的培训将允许学生更有效地交流通过自我监督的深度学习模式获得的知识。每年参与的学生将创建出现在各种应用程序中的常见术语或项的矢量表示，利用这些嵌入来改进预测模型，并适当地验证和记录这种方法的好处。在前5周，学生将接触到不同的嵌入策略和利用它们的机器学习应用程序，然后在最后5周过渡到他们个人研究工作的演示和故障排除。这个REU将帮助培养一支不仅擅长使用深度学习模型，而且能够扩展其功能的学生队伍。此外，该项目将培训来自研究资源有限的大学合作伙伴的不同类型的学生，以便与卡内基R1联邦指定的拉美裔服务机构成功的团队中的不同学生接触。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This three-year Research Experiences for Undergraduates (REU) site at the University of North Texas will support 10 students for 10 weeks each summer and train them to build modern artificial intelligence (AI) systems capable of communicating their acquired knowledge across domains. State of the art AI systems for challenging goals such as visual object recognition, speech recognition, and understanding natural language require extensive training with much of the learned knowledge locked into the structure of the built AI system. However, students in this program will focus on creating and leveraging highly-trained AI systems to re-represent information so the nuanced understanding learned by these systems is preserved and readily accessible for other applications. This REU brings together an interdisciplinary team to support projects that showcase the benefits of AI systems built to provide and receive inferred knowledge directly. Early in the program each student will identify an application domain and an external advisor to work with among a number of presented options. Students will also participate in a long-standing, unique AI summer research program integrating current university students and external REU students to facilitate collaboration across departments and student expertise.Specifically, the training in this REU will allow students to more efficiently communicate the knowledge acquired by self-supervised deep learning models. The participating students each year will create vector representations of common terms or items that appear across various applications, utilize those embeddings to improve prediction models, and properly validate and document the benefit of this approach. For the first 5 weeks, the students will be exposed to different embedding strategies and machine learning applications utilizing them, then transition to demonstrations and trouble-shooting of their individual research efforts in the last 5 weeks. This REU will help prepare a workforce of students not only adept at using deep learning models but extending their functionality. Additionally, this project will train a diverse range of students from college partners with limited research resources to engage with diverse students in successful teams with a Carnegie R1 federally designated Hispanic-serving institution.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.