SHF:Small:Scalable Spiking Neural Network Enabled by Probabilistic and Non-Volatile Synapses

集成电路(IC)技术在过去几十年的规模使计算速度和能效得到了显著的提高，我们今天手中的手机计算就是明证，不久前我们手中的手机计算还相当于房间大小的超级计算机。但当我们达到在纳米级扩展到更小特征尺寸的基本物理极限时，可以存储大量数据，并且超高速电路可以处理数据，因此现在整体系统性能受到在存储器和处理器之间传输数据时形成的瓶颈的限制。出于这个原因，替代计算模型，特别是那些基于大脑启发(神经形态)计算的模型，最近重新成为某些类型的计算应用和问题的潜在的新计算范例。这需要设备、电路和计算体系结构的进步，以及教育平台的创建，使未来的工程师和计算机科学家能够进步和利用它们。这项拟议工作的核心是使用一种新的磁性设备，它与传统集成电路技术相结合，以实现可扩展和高能效的神经形态计算机芯片。该设计将进行优化，以处理需要极高能效实施的“大数据”问题，例如为医学成像应用程序实时处理视频流。由于各种原因，这样的实现具有挑战性，最明显的是存储大量人工突触权重的值，以及计算做出人工神经元放电决策所需的随机数的有效方法。卡内基梅隆大学的研究人员正专注于解决拟议工作中的这两个关键挑战。

Scaling of integrated circuit (IC) technology for the past few decades has enabled remarkable advancement of computing speed and power efficiency, as evidenced by the cellphone computing that we hold in our hands today that would have corresponded to room-size super computers not long ago. But as we reach fundamental physical limits for scaling to smaller feature sizes at nanometer scale, massive amounts of data can be stored, and super-fast circuits can process the data, such that now the overall system performance is limited by the bottleneck that forms with transferring the data between the memory and the processor. For this reason, alternative computation models, particularly those based on brain-inspired (neuromorphic) computation, have recently resurged as a potential new computing paradigm for certain classes of computing applications and problems. This requires advancements in devices, circuits and computing architectures, along with creation of the education platform that will allow future engineers and computer scientists to advance and exploit them.At the core of this proposed work is the use of a novel magnetic device that is combined with traditional integrated circuit technology to enable a scalable and power efficient neuromorphic computer chip. The design will be optimized to handle "big data" problems that require extremely power efficient implementations, such as real-time processing of a video stream for a medical imaging application. Such implementations are challenging for various reasons, most notably the storing of values for a large number of artificial synapse weights, and efficient methods to compute random numbers that are needed to make artificial neuron spiking decisions. Carnegie Mellon researchers are focused on addressing these two key challenges in the proposed work.