Power-Efficient Long Short-term Memory

如果人工智能可以嵌入到小型的自供电传感器中，它将在以下场景中发挥作用。在精准农业中，这种传感器可以预测霜冻，以避免作物受损;监测当地鸟类种群中有害物种与有益物种;评估土壤条件，以实现有效灌溉。在智能工厂中，这种传感器可用于预测性维护，从而最大限度地减少因未检测到的设备故障和过于频繁的定期维修而导致的昂贵停机时间。在消费者医疗保健中，这种传感器可以佩戴在身体上，以检测睡眠呼吸暂停，跟踪心理健康状态或测量血压，所有这些都没有笨重，突兀的电池或频繁充电的中断。不幸的是，即使在中等复杂的应用中，人工智能也需要比自供电传感器提供更多的电力。为了解决这个问题，我们建议开发一种新型的人工智能，它不需要比自供电传感器更多的功率来运行。研究的目标是设计、实现和评估一种模拟长短期记忆（LSTM），它的功率效率是最先进的16倍。我们将通过以下方法实现这种功率效率：与现有技术相比，我们将使用（1）更少的输入和（2）更少的操作。我们的方法与传统的功耗降低策略（如内存计算、权重量化、知识蒸馏或网络修剪）兼容。此外，我们的方法是强大的模拟失配。我们将在关键字定位任务中演示我们的方法，尽管相同的原理可以扩展到其他应用程序。拟议活动的智力意义在于，它将通过开发一种新的LSTM范式来推进无处不在的传感器领域，该范式的能效比最先进的高16倍。由此产生的节能无处不在的传感器将提高不同领域的能力，该奖项反映了NSF的法定使命，并被认为值得支持通过使用基金会的知识价值和更广泛的影响审查标准进行评估。

If artificial intelligence could be embedded into small, self-powered sensors, it would be useful in the following scenarios. In precision agriculture, such sensors could forecast frost to avoid crop damage; monitor the local bird population for harmful versus beneficial species; and assess soil conditions for efficient irrigation. In smart factories, such sensors could be used for predictive maintenance, thus minimizing the costly downtime that is caused both by undetected equipment failure and by overly-frequent, scheduled servicing. In consumer healthcare, such sensors could be worn on the body to detect sleep apnea, track mental health state or measure blood pressure, all without bulky, obtrusive batteries or the disruption of frequent recharging. Unfortunately, even in moderately complex applications, artificial intelligence requires more power than a self-powered sensor can provide. To address this problem, we propose to develop a new type of artificial intelligence that requires no more power to run than is available in a self-powered sensor.The goal of the research is to design, implement and evaluate an analog long short-term memory (LSTM) that is 16 times more power-efficient than the state-of-the-art. We will achieve this power efficiency with the following methods: we will use (1) fewer inputs and (2) fewer operations than the state-of-the-art. Our approach is compatible with conventional power reduction strategies like compute-in-memory, weight quantization, knowledge distillation or network pruning. Further, our approach is robust to analog mismatch. We will demonstrate our approach in a keyword spotting task, although the same principles can be extended to other applications. The intellectual significance of the proposed activity is that it will advance the field of ubiquitous sensors by developing a new LSTM paradigm that is 16 times more power efficient than the state-of-the-art. The resulting energy-efficient ubiquitous sensors will increase capabilities across different fields, from industrial to agricultural and consumer healthcare applications.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.