EFRI BRAID: Scalable-Learning Neuromorphics

人工智能领域的最新进展表明，通往更高智能的大脑启发系统的途径是通过显著增加网络规模。然而，这种相当进化的方法面临着严峻的挑战。训练用于自然语言处理的最大的机器学习模型需要几个月的数据中心规模的计算，换句话说，需要巨大的能量、时间和成本。例如，由于算法和硬件的改进，改进的余地似乎有限，最重要的是，进一步的进步不能再由半导体技术的扩展来推动。此外，最先进的模型依赖于具有大量训练数据的离线训练，即不具备持续实时学习的能力。这些挑战自然会引起人们对生物神经网络的更多关注，生物神经网络是在非常节能的“大脑”硬件上运行的卓越、敏捷和自适应智能的活生生的证明。刺激神经网络理论的最新发展提供了令人兴奋的机会，这是生物学上最合理的模型，以及用密集的新兴记忆设备实现的神经形态回路。该项目旨在利用这些进步，解决最紧迫的挑战，为具有实际有用（鲁棒、快速、廉价）学习能力的人脑级神经形态系统开发可扩展的算法和硬件。该项目将使神经形态系统在许多实际应用中具有直接重要性，包括自主机器人和车辆，以及生物医学，包括便携式和个人医疗设备。此外，拟议研究的广泛算法到系统的性质为高中生，本科生和研究生探索新颖的研究和接触新兴的神经形态计算领域提供了有吸引力的机会。拟议的项目将通过利用参与大学的项目开展外展活动，特别注重吸引少数民族学生。我们研究的关键特征是硬件友好的局部学习算法，持续在线“一次”学习的框架，以及内存计算硬件电路的变化容忍。具体来说，在算法方面，我们将重点放在具有生物学上合理的尖峰频率适应神经元的循环尖峰神经网络上。我们将以团队成员最近提出的局部学习算法为基础，通过突触可塑性促进更长时间尺度的学习。这些算法将进一步扩展到支持持续学习，并使用神经结构搜索技术与硬件协同优化。在硬件方面，重点是利用模拟内存计算的混合神经形态电路。关键的硬件挑战，如网络复杂性的扩展和鲁棒原位学习的实现，将通过利用超高密度的交叉点设备实现学习算法的固定值权重和具有短期和长期可塑性的新型耐变记忆突触来解决。算法和硬件电路将整体集成到学习到学习的尖峰神经网络框架中。这种框架使两种学习成为可能——一种是模仿发展性学习的缓慢增量式学习，另一种是利用网络动态的快速“一次性”学习。该奖项反映了美国国家科学基金会的法定使命，并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，被认为值得支持。

Recent advances in the field of artificial intelligence suggest that the route toward higher-intelligence brain-inspired systems is via significantly increasing network size. However, such a rather evolutionary approach faces dire challenges. Training the largest machine learning models for natural language processing requires months of data-center-scale computations, in other words, enormous energy, time, and cost. Reserves for improvements, e.g., due to the refinement of algorithms and hardware, seem limited, and most importantly, further advances could no longer be fueled by semiconductor technology scaling. Additionally, state-of-the-art models rely on offline training with vast amount of training data, i.e., are not capable of continual real-time learning. Such challenges naturally bring more attention to the biological neural networks, which are living proof of superior, agile, and adaptive intelligence running on very energy-efficient “brain” hardware. Exciting opportunities are presented by recent developments in the theory of spiking neural networks, the most biologically plausible models, and neuromorphic circuits implemented with dense emerging memory devices. The proposed project aims to capitalize on these advances and address the most pressing challenges to develop scalable algorithms and hardware for human-brain-scale neuromorphic systems with practically useful (robust, fast, inexpensive) learning capabilities. The project will enable neuromorphic systems of immediate importance for many practical applications, including autonomous robots and vehicles, and biomedicine, including portable and personal medical devices. Furthermore, the broad algorithm-to-system nature of the proposed research provides attractive opportunities for high-school, undergraduate, and graduate students to explore novel research and get exposed to the emerging field of neuromorphic computing. The proposed project will pursue outreach activities by leveraging programs at participating universities, with a particular focus on attracting minority students.The key features of our research are hardware-friendly local learning algorithms, a framework for continual online “one-shot” learning, and variation-tolerant in-memory computing hardware circuits. Specifically, on the algorithmic front, we focus on recurrent spiking neural networks with biologically-plausible spike frequency adaptation neurons. We will build on local learning algorithms recently proposed by our team members that facilitate learning over longer time scales via synaptic plasticity. These algorithms will be further extended to support continual learning and co-optimized with the hardware using neural architecture search techniques. On the hardware front, the focus is on hybrid neuromorphic circuits that take advantage of analog in-memory computing. Critical hardware challenges, such as the scaling of network complexity and implementation of robust in-situ learning, will be addressed by utilizing ultra-high-density crosspoint devices implementing fixed-value weights of the learning algorithms and novel variation-tolerant memristive synapses featuring both short-term and long-term plasticity. Algorithms and hardware circuits will be holistically integrated into the learning-to-learn spiking neural network framework. Such a framework enables two kinds of learning – a slow incremental one mimicking developmental learning and fast “one-shot” learning utilizing network dynamics.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.