Near-sensor Computing Using Low-cost Stochastic Circuits

许多新兴的计算系统都遵循“感觉群”模式，其中传感器获取的数据与对这些数据的计算紧密地交织在一起。在环境监测、偏远地区监测和可植入设备的健康数据跟踪等应用中，传感器节点的资源极其有限。它们不能包括高性能微处理器来运行处理采集到的原始传感器数据的软件，这表明需要通过无线链路将这些数据传输到功能强大的“基础设施核心”。近传感器计算是一个概念，旨在避免在相对缓慢、耗电、潜在不可靠和不安全的通信通道上传输原始数据。这是通过专用硬件块实现的，这些硬件块以资源高效的方式直接在源处处理传感器数据。近传感器计算对于涉及分类的应用特别有前途，因为要分析的数据量是巨大的，并且通过传感器节点的计算取代它们的传输所获得的收益特别高。本项目旨在发展方法，以实现低成本和节能的硬件电路近传感器计算遵循随机计算范式。随机计算为复杂函数提供了极其紧凑、容错和低功耗的实现，但代价是更长的计算时间和一定程度的不准确性。这使得随机电路（SCs）对近传感器计算特别有吸引力，在近传感器计算中，处理的传感器数据无论如何都是不准确的，而且计算往往不经常发生。这个项目的一个特别重点将是用于分类任务的神经网络（nn）的SC实现，从轻量级nn到用于深度学习的完全成熟的卷积nn。该项目将通过随机电路寻找神经网络的有效表示。处理大型和复杂神经网络的所需能力将需要在随机计算的理论基础上取得各种进展，包括更好地理解相关性，开发SCs的随机数生成理论，以及研究混合随机二进制体系结构。这些理论发现与所开发的神经网络结构将产生一种综合和优化方法。为此，该项目将分为六个任务和两个密切相关的领域，一个侧重于理论，一个侧重于应用。此外，我们将继续与密歇根大学安娜堡分校的John P. Hayes教授进行长期而富有成效的合作，他是SC领域的领先科学家之一。该项目的研究有可能将物联网系统的功能提升到一个新的水平，将智能引入微型设备，用于随机电路的优势：小尺寸、低功耗、容错性和生物相容性最为明显的应用。

Many emerging computing systems follow the “sensory swarm” paradigm, where acquisition of data by sensors is tightly intertwined with computations on these data. In applications like environmental monitoring, surveillance of remote areas and health data tracking by implantable devices, the sensor nodes are extremely resource-restricted. They cannot include high-performance microprocessors to run software that processes the acquired raw sensor data, suggesting the need to transmit these data to a powerful “infrastructural core” via wireless links. Near-sensor computing is a concept that aims at avoiding the transmission of raw data over relatively slow, power-hungry and potentially unreliable and insecure communication channels. This is achieved by dedicated hardware blocks that process sensor data directly at their source in a resource-efficient manner. Near-sensor computing is particularly promising for applications that involve classification, because the amount of data to be analyzed is immense and the gain from replacing their transmission by a computation at the sensor node is particularly high.This project aims at developing methods to realize low-cost and power-efficient hardware circuits for near-sensor computing following the Stochastic Computing paradigm. Stochastic computing provides extremely compact, error-tolerant and low-power implementations of complex functions, but at the expense of longer computation times and some degree of inaccuracy. This makes stochastic circuits (SCs) especially attractive for near-sensor computing, where the processed sensor data are inaccurate anyway and computations tend to occur infrequently. A special focus of this project will be the SC realization of neural networks (NNs) used for classification tasks, from lightweight NNs to fully-fledged convolutional NNs for deep learning. The project will search for effective representations of NNs by stochastic circuits. The desired ability to handle large and complex NNs will necessitate various advances in the theoretical foundations of stochastic computing, including a better understanding of correlations, developing a theory of random number generation for SCs, and investigating hybrid stochastic-binary architectures. The theoretical findings together with the NN structures developed will yield a synthesis and optimization methodology. To this end, the project will be organized in six tasks and two closely intertwined areas, one focusing on theory and one on applications. Moreover, we will continue our longstanding and fruitful collaboration with Prof. John P. Hayes of University of Michigan, Ann Arbor, one of the leading scientists in the SC domain. The research in this project has the potential to bring the capabilities of IoT systems to the next level by introducing intelligence into tiny devices for applications where the advantages of stochastic circuits: small size, low power consumption, error tolerance and bio-compatibility, are most pronounced.