SHF Small: Hierarchical Unsupervised Inference Using Robust Neuromorphic Computation

人类大脑包含大约1000亿个神经元。每个处理器的工作频率约为150赫兹，比现代数字处理器慢得多，这表明大脑的计算能力源于其大规模并行架构，而不是纯粹的处理速度。深度机器学习（DML）最近作为一种很有前途的框架出现，用于模仿大脑的信息表示能力。受神经生物学发现的启发，深度学习系统的隐藏层对复杂输入的分层分布表示进行编码。然而，传统处理器的高度并行架构和串行结构之间的根本不匹配限制了基于软件的DML系统的可伸缩性。通过充分利用单个晶体管的计算能力，模拟神经形态电路实现了比数字技术更高的密度和能量效率。本研究的目标是使用神经形态模拟计算元素来实现可扩展的深度学习系统。该研究为大规模深度机器学习系统的设计和使用提供了革命性的潜力，并将其应用于现实世界中复杂的高维模式识别问题。特别是，实现这种系统的浮门电路的研究有望在神经形态工程领域产生广泛的影响。紧凑、节能的学习系统的长期影响将是深远的，从能够从自己的观察中学习的微型健康监测传感器到多媒体数据的大规模自主处理，都将为社会带来益处。该项目的跨学科性质将为研究生和本科生提供一个在不同专业交叉点工作的机会，并提供一个独特的平台来研究复杂生物启发系统的紧急特性。

The human brain contains roughly 100 billion neurons. Each operates at approximately 150 Hz, far slower than modern digital processors, suggesting that the brain's computational strength stems from its massively parallel architecture rather than sheer processing speed. Deep machine learning (DML) has recently emerged as a promising framework for mimicking the information representation capabilities of the brain. Inspired by discoveries in neurobiology, hidden layers of deep learning systems encode hierarchically distributed representations of complex inputs. However, the fundamental mismatch between a highly parallel architecture and the serial structure of conventional processors limits the scalability of software-based DML systems. By fully leveraging the computational power of individual transistors, analog neuromorphic circuits achieve much greater density and energy efficiency than digital technology. The goal of this research is to use neuromorphic analog computational elements to enable scalable deep learning systems.The proposed research offers the potential to revolutionize the design and utilization of large-scale deep machine learning systems with applications to real-world, complex, high-dimensional pattern recognition problems. In particular, the study of floating-gate circuitry for realizing such systems is expected to have broad impact on the field of neuromorphic engineering. The long-range impact of compact, power-efficient learning systems will be profound, with benefits to society ranging from micro-scale health-monitoring sensors capable of learning from their own observations to large-scale autonomous processing of multimedia data. The interdisciplinary nature of the project will provide an opportunity for graduate and undergraduate students to work at the intersection of different specializations and provide a unique platform on which to study the emergent properties of complex biologically-inspired systems.