CNS Core: Small: Importance-Aware Compressive Inference for Efficient Embedded Vision

CNS 核心：小型：重要性感知压缩推理，实现高效嵌入式视觉

• 批准号: 2008151
• 负责人: Robert Dick
• 金额: $ 49.97万
• 依托单位: Regents of the University of Michigan - Ann Arbor
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2020
• 资助国家: 美国
• 起止时间: 2020-10-01 至 2023-09-30
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2008151&HistoricalAwards=false
• 关键词: CNS; Core; Small; Importance; Aware

This project explores the concept of Compressive Inference in the context of energy-efficient and low-latency machine learning and artificial intelligence (AI) applications. Compressive Inference is the biologically inspired idea of using highly heterogeneous, multi-round,feedback-controlled sampling and analysis of signals to minimize latency and energy consumption while maximizing inference accuracy, which stands in contrast to the commonly optimized but often less relevant objective of signal reconstruction accuracy. Methods of restructuring and compressing knowledge representations to improve efficiency are also being explored. The project focuses on applications and systems facing tight energy consumption and latency constraints, namely computer vision applications running on low-power embedded systems, although many of the ideas developed will have application in other domains, e.g., datacenter-based machine learning and AI applications.Based on preliminary results, it is likely that the project will enable order-of-magnitude improvements in machine learning and AI application inference latencies and energy consumptions, thereby enabling the deployment of sophisticated analysis techniques in applications where they were previously impractical, e.g., low-cost home security systems and agricultural sensing applications, as well as applications where sophisticated analysis was detrimentally resource and power hungry, e.g., autonomous driving and wearable vision-based assistants. The resulting improvement in efficiency will enable local, on-device learning, thereby making it possible for machine learning and AI systems to adapt to their environments, thereby reducing the tendency to perform poorly on data dissimilar to samples in centralized training datasets. The project also has an educational component, in which students with a broad range of backgrounds will learn about state-of-the-art approaches to machine learning, AI, and embedded system design.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.

这个项目在节能和低延迟的机器学习和人工智能(AI)应用的背景下探索压缩推理的概念。压缩推理是一种生物启发的思想，它使用高度不同的、多轮的、反馈控制的信号采样和分析来最小化延迟和能量消耗，同时最大化推理精度，这与通常优化但通常不太相关的信号重建精度的目标形成了鲜明对比。还在探索重组和压缩知识表示以提高效率的方法。该项目专注于面临严格的能耗和延迟限制的应用程序和系统，即在低功率嵌入式系统上运行的计算机视觉应用程序，尽管所开发的许多想法将在其他领域中应用，例如基于数据中心的机器学习和人工智能应用程序。基于初步结果，该项目很可能使机器学习和人工智能应用程序推理延迟和能量消耗得到数量级的改进，从而使复杂的分析技术能够部署在以前不切实际的应用程序中，例如低成本的家庭安全系统和农业传感应用程序，以及复杂分析有害于资源和耗电的应用程序，例如自动驾驶和可穿戴的基于视觉的助手。由此产生的效率提高将使本地、设备上的学习成为可能，从而使机器学习和人工智能系统能够适应其环境，从而减少在中央训练数据集中对不同于样本的数据表现不佳的倾向。该项目还包括一个教育部分，在其中，具有广泛背景的学生将学习机器学习、人工智能和嵌入式系统设计的最先进方法。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。

项目成果

How Capable Can a Transformer Become? A Study on Synthetic, Interpretable Tasks

变压器能变得多强大？

• 发表时间: 2023
• 期刊: Proc. NeurIPS Wkshp. on Symmetry and Geometry in Neural Representations
• 作者: Ramesh, Rahul;Mikail Khona;Robert P. Dick;Hidenori Tanaka;Ekdeep Singh Lubana.
• 通讯作者: Ekdeep Singh Lubana.

Compositional Abilities Emerge Multiplicatively: Exploring Diffusion Models on a Synthetic Task

• DOI: 10.48550/arxiv.2310.09336
• 发表时间: 2023-10
• 期刊: ArXiv
• 作者: Maya Okawa;Ekdeep Singh Lubana;Robert P. Dick;Hidenori Tanaka

Orchestra: Unsupervised Federated Learning via Globally Consistent Clustering

• DOI: 10.48550/arxiv.2205.11506
• 发表时间: 2022-05
• 作者: Ekdeep Singh Lubana;Chi Ian Tang;F. Kawsar;R. Dick;Akhil Mathur

Beyond BatchNorm: Towards a Unified Understanding of Normalization in Deep Learning

• 发表时间: 2021-06
• 作者: Ekdeep Singh Lubana;R. Dick;Hidenori Tanaka

How do quadratic regularizers prevent catastrophic forgetting: The role of interpolation

二次正则化器如何防止灾难性遗忘：插值的作用

• 发表时间: 2022
• 期刊: Proc. Conf. on Lifelong Learning Agents
• 作者: Lubana, Ekdeep Singh;Puja Trivedi;Danai Koutra;Robert Dick.
• 通讯作者: Robert Dick.