UKRI-RCN: Exploiting the dynamics of self-timed machine learning hardware (ESTEEM)

如今，人们经常可以在媒体公司的面包车上看到这样的广告，比如《千兆抢夺你的生活》(Virgin Media)。类似的口号出现在IT公司的传单上，提供每秒Tera操作的数据分析。它们显示了无可否认的技术进步，尽管我们仍然很少看到每种能源的性能增长，例如每焦耳千兆位。然而，我们越来越多地不得不面对不断上涨的能源账单。随着将我们的智能扩展到我们日常生活的更广泛和更深入的需求稳步增长，信息和通信技术在提高智能能源效率方面的巨大挑战变得越来越明显。在这方面的一个重要作用属于旨在找到更好的机器学习和数据分类方法的研究，在这种方法中，执行学习中的关键操作的权力和时间都减少了。简而言之，降低功率相当于减少电子硬件的平均开关活动，而减少时间意味着确定学习动作达到足够质量状态的时刻。自时间硬件基于事件驱动原理，结合新的机器学习方法，基于有效的近似和布尔逻辑，而不是重算术，为这项研究提供了一个创新杠杆，并针对当前的技术状况产生了潜在的影响。该项目将研究通过自定时电路固有的时间和功率弹性来提高人工智能硬件的性能和能效的机会。该项目将为在微米和纳米尺度上构建具有机器学习(ML)能力的电子设备和系统的新设计方法奠定基础。这些设备将被广泛利用在许多边缘应用中，如环境传感器、交通监测器、可穿戴设备，以及潜在的可用作未来计算机系统构建块的商用ML增强型设备。能够以功率/能源效率实时运行的微纳分类器和决策者有望找到许多“轻量级”应用，因此优化(在延迟和能量方面)控制至关重要。这是一个用具有能量采集能力的电子笔识别手写字符的例子。给出了一个参考类别(例如，数字“5”)。然后，进行几次数字5的手写尝试。在所有这些尝试中，都进行了培训。然后给出另一个参考课，并对其进行类似的培训，以此类推。关键的要求是将花费的时间保持在有限的范围内，并将消耗的能量降至最低。训练应以可达到的最佳精度进行。在学习的速度、力量和准确性之间有几个权衡。该项目的成功将通过对有关自定时电路中机器学习动力学的关键研究问题的回答来衡量；例如，结合使用学习自动机和基于逻辑的推理的异步设计方法是否会达到给定机器学习问题的最小能量点。理论和设计方法上的项目成果将通过广泛的模拟、原型、IC制造和测试，并最终通过将新的硬件解决方案应用到具体的物联网应用中来验证。一个特别具有挑战性和突破性的验证将是开发和制造第一个使用柔性基板的异步机器学习集成电路。这项研究的实际影响将是设计智能嵌入式电子产品的方向和方法，这些电子产品将能够对从环境传感器、音频和图像信号以及快速移动消费品(FMCG)和使用柔性IC技术的智能包装获得的数据执行运行时分类。

People can often see these days adverts on media company vans like "Grab your life by the Gigabits" (Virgin Media). Similar slogans appear on IT company flyers offering data analysis at Tera operations per second. They show the undeniable progress in technology, though still rarely we see performance growth per energy, for example Gigabits per Joule. And yet we are increasingly having to face with rising energy bills. As appetites for extending our intelligence wider and deeper into our everyday life steadily grow the grand challenge of ICT in making intelligence energy efficient becomes more and more evident. A significant role in this belongs to the research that aims at finding better methods for machine learning and data classification where both power and time for performing key operations in learning are reduced. In simple terms reducing power amounts to reducing average switching activity of electronic hardware, while reducing time means determining the moments when the learning actions have reached the state of sufficient quality. Self-time hardware, which works on the event-driven principles, in combination with novel machine learning methods, based on efficient approximation and Boolean logic as opposed to heavy arithmetic, gives this research a lever of innovation and potential impact against the state of the art.This project will investigate opportunities for improving performance and energy efficiency in artificial intelligence hardware created by the inherent time and power elasticity of self-timed circuits. The project will lay foundation to a new design methodology for building electronic devices and systems with machine learning (ML) capabilities at the micro- and nano-scale granularity. Those devices will be widely leveraged in many at-the-edge applications such as environmental sensors, traffic monitors, wearables, as well potential commodity ML-enhanced devices that can be used as building blocks in computer systems of the future. Micro- and nanoclassifiers and decision makers that can operate in real-time with power/energy efficiency are expected to find many 'light-weight' applications, so optimal (in terms of latency and energy) control is crucial. Here is an example of handwritten character recognition by an electronic pen with energy-harvested power. A reference class is given (e.g., digit "5"). Then, a few attempts in handwriting of digit 5 are made. During all these attempts training is performed. Then another reference class is given, and similar training is performed on it, and so on. The key requirements are to keep time spent limited and consumed energy minimised. Training is to be done to the best of the achievable accuracy. There are several trade-offs involved between speed and power and accuracy of learning. The success of the project will be measured in terms of the answers to the key research questions about the dynamics of machine learning in self-timed circuits; for example, whether the asynchronous design approach combined with the use of learning automata and logic-based inference will reach minimum energy point for a given machine learning problem. The project outcomes in theory and design methodology will be validated by means of extensive simulations, prototyping, IC fabrication and testing, and, ultimately, via an embodiment of the new hardware solutions into a concrete IoT application. A particularly challenging and breaking through validation will be the development and fabrication of the first asynchronous machine learning integrated circuit using flexible substrates.The practical impact of this research will be in the directions and methods of designing intelligent embedded electronics that will be capable of performing run-time classification of data obtained from environmental sensors, audio and image signals, as well as fast moving consumer goods (FMCG) and smart packaging using flexible IC technology.