EFRI BRAID: Efficient Learning of Spatiotemporal Regularities in Humans and Machines through Temporal Scaffolding

在最小的能量预算下运行，人脑能够在不同的时间尺度上高效地处理大量的时间信息，因为它很快就学会了在新的环境中行动。相比之下，目前的人工智能模型不能有效地学习时间信息，难以实现终身学习--即在一生中不断学习新任务的能力--而且在资源有限的环境中也表现不佳。该项目旨在创建新的人工智能模型，通过利用大脑如何能够有效学习时间信息的理论启发的机制来克服这些限制。特别是，该项目是基于最近的一项名为“时间脚手架”的理论，该理论假设，在睡眠期间，大脑会以一种加速的方式重新激活醒来的经历，从而能够检测出嵌入这些经历的重要的时间模式。这个项目的目标是开发基于时间脚手架假设的自主机器，它可以快速适应，在不确定的情况下操作，并在资源受限的情况下在整个生命周期内进化。这种变革性的方法有可能应对主要的人工智能挑战，并在医疗保健、能源和国家安全方面找到应用。该小组的目标是通过在多个教育机构采取举措，促进更广泛地利用计算战略，强调跨学科培训和面向代表性不足的人群。该小组将在整个项目期间举办对价值敏感的讲习班和定期道德操守协商。除了技术目标，该团队的目标是为人工智能领域代表性不足的学生提供机会，培养一支具有竞争力的人工智能劳动力，以保持美国在STEM的技术领先地位。通过模仿人脑的学习方式，该团队试图创建高效的、终身学习的人工智能系统，该系统能够彻底改变各个行业，并造福于整个社会。时间脚手架假说为大脑高效学习时间信息的优越能力提供了一种新的解释。根据这一假设，离线期间的时间压缩记忆重放用于提取编码经验中的时间规律性。在时间脚手架假说的基础上，在本项目中，PI提出了一套基于资源高效的终身学习时空规则的机制，使用在线(“觉醒”)和离线(“睡眠”)时间段，他们打算在新的人类实验中验证这些机制，并将其纳入机器学习算法。这笔赠款在理论、模型和系统方面的进步将在多个领域得到应用。该项目的两个具体目标是：i)在时间支架假设的启发下，开发新的人工智能算法和体系结构，用于时空模式的有效学习；ii)扩展时间支架假设，包括支持终身学习的层次表示，并通过人体实验和计算研究验证模型的预测。通过这些目标，PI将开发支持在资源受限环境中部署的优化框架。此外，该项目将产生可扩展的深度神经网络和尖峰神经网络模型，其中包含时间压缩重放机制。这种方法预计将限制阻碍大多数当前记忆网络模型的灾难性干扰效应，并提高系统的终身学习能力。通过德克萨斯大学圣安东尼奥分校、罗切斯特大学和诺克斯维尔田纳西大学的多项倡议，包括成功的K-12合作伙伴关系和有针对性的体验式推广战略，将扩大培训和获得这些变革性计算战略的机会。该小组还将通过对价值有敏感认识的讲习班和定期道德操守协商，参与道德操守设计。项目设计工作还将为交叉人工智能领域中代表性不足的学生提供重要机会，并促进一支强大而有竞争力的人工智能劳动力队伍，以保持美国在STEM的技术领先地位。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。

Operating on a minimal energy budget, the human brain is able to efficiently process vast amounts of temporal information at different timescales as it quickly learns to act in new environments. By contrast, current AI models do not learn temporal information efficiently, struggle with lifelong learning - the ability to keep learning new tasks continuously throughout life - and also do not perform well in resource-constrained environments. This project aims to create new AI models that overcome these limitations by leveraging mechanisms inspired by theories for how the brain is able to efficiently learn temporal information. Particularly, the project is based on a recent theory, 'temporal scaffolding', which postulates that during sleep, the brain reactivates wake experiences in an accelerated manner to allow detecting important temporal patterns embedded in those experiences. The goal of this project is to develop autonomous machines, informed by the temporal scaffolding hypothesis, which can rapidly adapt, operate under uncertainty, and evolve throughout their lifespan despite resource constraints. This transformative approach has the potential to address major AI challenges and find applications in healthcare, energy, and national security. The team aims to promote broad access to the computational strategies through initiatives at multiple educational institutions, emphasizing cross-disciplinary training and outreach to underrepresented populations. The team will conduct value-sensitive workshops and regular ethics consultations throughout the project. Alongside the technical goals, the team aims to offer opportunities for underrepresented students in AI fields, fostering a competitive AI workforce to maintain US technological leadership in STEM. By emulating how the human brain learns, the team seeks to create efficient, lifelong learning AI systems capable of revolutionizing various industries and benefiting society as a whole.The Temporal Scaffolding Hypothesis provides a novel explanation for the brain’s superior ability to efficiently learn temporal information. According to this hypothesis, time-compressed memory replay during offline periods serves to extract temporal regularities within encoded experiences. Building on the temporal scaffolding hypothesis, in the present project the PIs propose a set of mechanisms underlying resource-efficient lifelong learning of spatiotemporal regularities employing online (“wake”) and offline (“sleep”) periods, which they intend to both verify in new human experiments and incorporate in machine learning algorithms. Advances in theory, models, and systems stemming from this grant will have applications in multiple domains. The two specific aims for this project are to: i) develop new AI algorithms and architectures, inspired by the temporal scaffolding hypothesis, for efficient learning of spatiotemporal patterns and ii) extend the temporal scaffolding hypothesis to include hierarchical representations that support lifelong learning and verify the predictions of the model through human experiments and computational investigations. Through these aims the PIs will develop optimization frameworks that support deployment in resource constrained environments. Moreover, this project will yield scalable deep neural network and spiking neural network models that incorporate temporally compressed replay mechanisms. This approach is expected to limit the catastrophic interference effects that hinder most current network models of memory and improve the system’s capacity for lifelong learning. Training and access to these transformative computational strategies will be broadened via multiple initiatives at the University of Texas at San Antonio, the University of Rochester, and the University of Tennessee, Knoxville, including successful K-12 partnerships and targeted experiential outreach strategies. The team will also engage in ethical design through value-sensitive workshops and regular ethics consultations. The project design efforts will also provide significant opportunities to underrepresented students in cross-cutting AI fields and promote a robust and competitive AI workforce that maintains US technological leadership in STEM.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.