Neuromorphic-Digital Intelligent Systems

未来的人工智能系统将包括不同架构类型和范式的多个神经网络。这些未来的系统可能包含多个自主AI组件，这些组件不断协作学习并相互通信，以基于数据/传感器输入的组合来获取知识。目前，大多数AI都是建立在基于晶体管的硬件上。因此，我们可能会根据我们目前对数字人工神经网络（ANN）的理解，开始设想未来的复杂智能系统，其中协作组件之间的通信将是数字化的并符合冯·诺依曼架构。随着神经形态技术成为现实，未来的人工智能系统可能会包括与成熟的数字世界结合使用的尖峰神经网络（SNN）。此外，随着混合系统的出现，神经形态计算的采用可能是渐进的，混合系统最大化了神经形态SNN方法的优势。（能量减少，神经处理不可或缺的记忆，通过时间稀疏性提高效率，高保真神经形态数据）和数字ANN替代品（丰富的数字数据，使用已建立的反向传播的现有深度学习算法，云提供的GPU和CPU，已建立的编程模型）。这种混合方法可能会继续存在，因为与生物智能不同，机器智能需要直接与数字（在线）和物理（真实的）世界进行交互。本研究项目将在软件层探索这些不同技术之间的差距。它将研究如何最佳地将它们联合收割机结合起来，例如通过学习不同类型的表示和学习范式之间的映射。

Future AI systems will comprise multiple neural networks of different architectural types and paradigms. These future systems may contain multiple autonomous AI components that continually learn collaboratively and communicate with each other to derive knowledge based on a combination of data/sensor inputs. Currently, most AI is built on transistor-based hardware. Therefore, we might begin by envisaging future complex intelligent systems based on our current understanding of digital artificial neural networks (ANNs), whereby communications between collaborating components would be digital and conform to Von Neumann architecture. As neuromorphic technologies become reality, future AI systems are likely to include spiking neural networks (SNNs) working in conjunction with the well-established digital world. Furthermore, the adoption of neuromorphic computing is likely to be incremental, with the emergence of hybrid systems that maximise the advantages of neuromorphic SNN approaches (energy reduction, memory integral to neural processing, efficiency through temporal sparsity, high fidelity neuromorphic data) and digital ANN alternatives (abundance of digital data, existing deep learning algorithms using established back propagation, cloud provisioned GPUs and CPUs, established programming models). This hybrid approach is likely to remain because, unlike biological intelligence, machine intelligence will need to interact directly with both the digital (online) and physical (real) worlds. This research project will explore the gap between these different technologies at the software layer. It will investigate ways to optimally combine them, for example by learning mappings between different types of representation and learning paradigms.