HAINEM: Hardware Software Co-Design Solutions for Ultra-Efficient Artificial Intelligence and Neuromorphic Systems by Using Emerging Memory Technologies

传统的计算平台，特别是基于冯-诺依曼的计算架构，严重地受到处理单元和存储器之间的有限数据带宽的影响，这使得它们不适合大数据处理和实现具有大量参数集的最先进的机器学习（ML）算法。HAINEM计划将提出硬件软件协同设计解决方案，为人工智能（AI）应用开发超高效的专业计算平台。在这个项目中，我们通过配置内存和处理来部署新兴电阻开关（RS）存储器技术设备的内在行为，以实现内存计算（CIM）并解决称为内存墙的数据通信问题。我们将超越冯·诺依曼应用处理单元（APU），通过利用硬件软件协同设计解决方案来加速ML算法并减轻硬件非理想性。HAINEM将研究新的网络，并将现有的机器学习问题重新构建为更兼容的算法，用于拟议的APU。 HAINEM的第一个里程碑是开发边缘ML硬件，通过在基于RS的交叉开关加速器中获得完全并行模拟VMM的优势，提高计算效率（定义为TOPs/mm 2）和能效（定义为TOPS/W）。它的目标是开发下一代机器学习硬件，可以嵌入到前端传感器网络附近的边缘，直接处理信号，而无需将其发送到云端。该计划还将专注于神经形态计算概念，灵感来自最有效的计算系统，人类大脑，通过尖峰神经网络（SNN）结构进一步提高效率，实现更并行，低功耗和强大的计算平台。在HAINEM中，我们展示了基于RS的系统如何模仿SNN中的功能基元，并可以复制突触和神经元模型，如学习行为作为我们大脑中的生物突触，如长时程抑郁（LTD）和长时程增强（LTP）。我们还将探索不同的新兴RS记忆设备作为突触连接，并将在这些突触设备上测试各种生物启发的学习机制，如基于对和基于三重峰的尖峰时间依赖可塑性（STDP）。HAINEM还尝试通过使用混合信号电路设计概念来开发更有效的基于RS的神经元电路。最后，我们的目标是在芯片上设计功能齐全的神经形态架构，并将展示近传感器计算应用的实现，如EEG癫痫发作检测，基于EEG的大脑状态识别，视网膜数据处理。HAINEM将是一个重要的研究计划，帮助加拿大和多伦多成为人工智能和神经形态计算硬件实现的主要研发中心，与安大略在基于软件的神经网络方面的世界级专业知识形成强大的协同作用。

Conventional computing platforms, specially von-Neumann-based computing architectures, are seriously suffering from the limited data bandwidth between processing units and memory which makes them incompatible for big data processing and implementing state-of-the-art machine learning (ML) algorithms with extensive large sets of parameters. The HAINEM program will propose the hardware software co-design solutions to develop ultra-efficient specialized computing platforms for artificial intelligence (AI) applications. In this program, we deploy the intrinsic behaviour of emerging resistive switching (RS) memory technology devices by collocating memory and processing to enable compute-in-memory (CIM) and resolve the data communication problem known as memory wall. We will develop beyond von-Neumann application processing units (APUs) to accelerate ML algorithms and mitigating the hardware non-idealities by leveraging hardware software co-design solutions. HAINEM will investigate new networks and reframe existing machine learning problems into more compatible algorithms for the proposed APUs. HAINEM's very first milestone is to develop the edge ML hardware to improve computational efficiency (defined as TOPs/mm2) and energy efficiency (defined as TOPS/W) by getting the advantage from fully parallel analog VMM in RS-based crossbar accelerators. It aims to develop next generation ML hardware which could be embedded into the edge, near the front-end sensors networks, to process signals directly without sending them toward the cloud. This program will also focus on neuromorphic computing concepts inspired from the most efficient computing system, human brain, to further enhance the efficiency toward a more parallel, low-power, and robust computing platform through spiking neural network (SNN) structure. In HAINEM, we show how RS-based systems mimic functional primitives in SNN and can reproduce synaptic and neuronal models like learning behaviours as biological synapses in our brain such as long term depression (LTD) and long term potentiation (LTP). We will also explore different emerging RS memory devices as synaptic connections and will test various bio-inspired learning mechanisms over these synaptic devices like pair-based and triplet-based spike-time-dependant plasticity (STDP). HAINEM also try to develop more efficient RS based neuronal circuit by using mixed-signal circuit design concepts. Finally, our goal is to design fully functional neuromorphic architectures on chip and proof of concept will show the implementation of near-sensor computing applications such as EEG seizure detection, EEG-based brain state identification, retina data processing. HAINEM will be an important research program which helps Canada and Toronto to become a major research and development hub for hardware implementation for AI and neuromorphic computing in strong synergy with Ontario's world-class expertise in software-based neural networks.