EFRI BRAID: Unsupervised Continual Learning with Hierarchical Timescales and Plasticity Mechanisms

EFRI BRAID：具有分层时间尺度和可塑性机制的无监督持续学习

• 批准号: 2223793
• 负责人: Gianfranco Doretto
• 金额: $ 200万
• 依托单位: West Virginia University Research Corporation
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2022
• 资助国家: 美国
• 起止时间: 2022-10-01 至 2026-09-30
• 项目状态: 未结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2223793&HistoricalAwards=false
• 关键词: EFRI; BRAID; Unsupervised; Continual; Learning

Humans and animals can easily adapt to their environment with limited information. They sense the world around them and continuously adapt their behavior to the current situation by changing the “configuration” of their nervous system, a phenomenon called plasticity. Though this ability seems natural to humans, it is very difficult to achieve in software or hardware systems. In addition, current continuous learning methods are trained under unrealistic conditions and require supervision. This project aims to understand how to endow autonomous agents, such as robots, with the adaptability and resiliency of biology. Biological plasticity in weakly electric fish will guide engineering of new machine learning algorithms. These algorithms will enable autonomous agents to continuously sense and adapt to their environment without interrupting operations for manual training. This interdisciplinary project is integrated with a range of outreach activities involving local high schools and undergraduate students. Workshops and demonstrations on biology-inspired machine learning will be organized, aimed at spurring interest of rural students in coding and robotics.A grand challenge in artificial intelligence (AI) is how to achieve unsupervised continual learning in the open world. Current methods used in AI and machine learning operate with single-modality data, collected and consumed in controlled conditions, typically in a supervised manner. However, biological systems achieve lifelong learning by processing streams of multisensory data that continuously shape their neural networks (plasticity) while retaining previous knowledge (stability). This dynamic adaptation operates unsupervised, on a range of timescales and rules. The project will study those principles observed in the cerebellar feedback pathways of electric fish, which are responsible for driving plasticity, enabling adaptation of its function at different timescales and learning and forgetting at multiple speeds. This will enable the translational development of novel paradigms in continual learning that will support new levels of resiliency and lifelong learning in real-time autonomous systems in the open world. To achieve this goal the project will overcome some key technical hurdles, e.g., in enabling 1) data efficiency in processing inputs continuously as time-variant, potentially correlated, data streams in a fully unsupervised manner; 2) flexibility to learn and forget at different speeds; 3) generation of suitable internal representations from multiple modalities to improve autonomous resilience. This project is jointly funded by the Emerging Frontiers in Research and Innovation Brain-Inspired Dynamics for Engineering Energy-Efficient Circuits and Artificial Intelligence Program (BRAID) and the Established Program to Stimulate Competitive Research (EPSCoR).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.

人类和动物可以很容易地适应他们的环境有限的信息。他们感知周围的世界，并通过改变神经系统的“配置”来不断调整自己的行为以适应当前的情况，这种现象称为可塑性。 虽然这种能力对人类来说似乎是自然的，但在软件或硬件系统中很难实现。此外，目前的持续学习方法是在不现实的条件下训练的，需要监督。该项目旨在了解如何赋予自主代理，如机器人，生物学的适应性和弹性。弱电鱼的生物可塑性将指导新机器学习算法的工程设计。这些算法将使自主代理能够持续感知和适应环境，而无需中断操作进行手动培训。这一跨学科项目与一系列涉及当地高中和本科生的外联活动相结合。将组织关于生物启发的机器学习的研讨会和演示，旨在激发农村学生对编码和机器人技术的兴趣。人工智能（AI）的一个重大挑战是如何在开放世界中实现无监督的持续学习。目前人工智能和机器学习中使用的方法使用单模态数据，这些数据在受控条件下收集和消费，通常以监督的方式进行。然而，生物系统通过处理多感官数据流来实现终身学习，这些数据流不断塑造其神经网络（可塑性），同时保留先前的知识（稳定性）。这种动态适应在一系列时间尺度和规则上无监督地运行。该项目将研究在电鱼小脑反馈通路中观察到的那些原理，这些原理负责驱动可塑性，使其功能在不同的时间尺度上适应，并以多种速度学习和遗忘。这将使新的范式在持续学习中的转化发展成为可能，这将支持开放世界中实时自主系统的新水平的弹性和终身学习。为了实现这一目标，该项目将克服一些关键的技术障碍，例如，在实现1）以完全无监督的方式连续处理输入作为时变的、潜在相关的数据流的数据效率; 2）以不同速度学习和遗忘的灵活性; 3）从多个模态生成合适的内部表示以提高自主弹性。该项目由新兴前沿研究和创新脑启发动力学工程节能电路和人工智能计划（BRAID）和刺激竞争研究的既定计划（EPSCoR）共同资助。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

DNA: Deformable Neural Articulations Network for Template-free Dynamic 3D Human Reconstruction from Monocular RGB-D Video

DNA：可变形神经关节网络，用于从单目 RGB-D 视频进行无模板动态 3D 人体重建

• DOI: 10.1109/cvprw59228.2023.00375
• 发表时间: 2023
• 期刊: IEEE
• 作者: Vo, Khoa;Pham, Trong-Thang;Yamazaki, Kashu;Tran, Minh;Le, Ngan
• 通讯作者: Le, Ngan

A perspective on the neuromorphic control of legged locomotion in past, present, and future insect-like robots

• DOI: 10.1088/2634-4386/acc04f
• 发表时间: 2023-03
• 期刊: Neuromorphic Computing and Engineering
• 作者: N. Szczecinski;C. Goldsmith;W. Nourse;R. Quinn

More Synergy, Less Redundancy: Exploiting Joint Mutual Information for Self-Supervised Learning

• DOI: 10.1109/icip49359.2023.10222547
• 发表时间: 2023-07
• 期刊: 2023 IEEE International Conference on Image Processing (ICIP)
• 作者: S. Mohamadi;Gianfranco Doretto;Don Adjeroh

CLIP-TSA: Clip-Assisted Temporal Self-Attention for Weakly-Supervised Video Anomaly Detection

• DOI: 10.1109/icip49359.2023.10222289
• 发表时间: 2022-12
• 期刊: 2023 IEEE International Conference on Image Processing (ICIP)
• 作者: Hyekang Joo;Khoa T. Vo;Kashu Yamazaki;Ngan T. H. Le

A Robust Likelihood Model for Novelty Detection

• DOI: 10.48550/arxiv.2306.03331
• 发表时间: 2023-06
• 期刊: ArXiv
• 作者: Ranya Almohsen;Shivang Patel;Don Adjeroh;Gianfranco Doretto