SPIKEPro - SPIKING PHOTONIC-ELECTRONIC IC FOR QUICK AND EFFICIENT PROCESSING

人工智能技术的快速发展带来了强大的模型和算法，这些模型和算法彻底改变了所有科学和技术领域的许多应用。在人工神经网络中进行的深度学习产生了处理数据的新方法，从而产生了具有令人印象深刻的功能和优势的复杂系统。然而，传统的计算硬件在能源效率和速度方面已经达到了极限。需要一种新的计算硬件方法。新型的受大脑启发或神经形态芯片与受生物启发的脉冲神经网络一起工作，因为它们承诺以高效的方式处理数据而受到关注。在电子或光子硬件中分别开发这种神经形态系统的重要研究工作已经投入，每种系统都有其缺点和局限性。SPIKEPro提出了一个科学走向技术的突破，将低能量的电子和光子神经元结合到集成电路上的联合脉冲神经网络中。SPIKEPro的芯片集成方法基于通用技术平台，通过非易失性突触权重将超快激光光学神经元与高效电尖峰二极管连接起来。这使得我们能够同时利用电子和光子学的优势，提供超越现有实现的高效高速snn。除了减少网络中每个尖峰的能量消耗外，SPIKEPro还将开发新的学习策略和算法，能够在减少突触连接数量的情况下工作。这将通过利用电和光子尖峰器件的硬件参数来实现。SPIKEPro的成果将产生持久的经济、社会和科学影响。该项目将为边缘计算、传感器数据处理、高速控制和计算神经科学等不同领域带来超快速高效的神经形态硬件。

Rapid advances in artificial intelligence technologies have led to powerful models and algorithms that have revolutionized many applications across all fields of science and technology. Deep learning performed within artificial neural networks has yielded new ways to process data, leading to sophisticated systems with impressive functionality and benefits. However, conventional computing hardware is reaching its limits in terms of energy efficiency and speed. A new approach to computing hardware is needed. Novel brain inspired or neuromorphic chips working with biologically-inspired spiking neural networks have gained attention as they promise highly efficient ways to process data. Important research effort has been dedicated to develop such neuromorphic systems in electronic or photonic hardware separately, each with its drawbacks and limitations. SPIKEPro proposes a science-towards-technology breakthrough by combining low-energy electrical and photonic neurons into a joint spiking neural network on an integrated circuit. SPIKEPro’s chip integration approach is based on a common technology platform, connecting ultrafast laser optical neurons with efficient electrical spiking diodes through non-volatile synaptic weights. This enables to simultaneously capitalise on the advantages of both electronics and photonics to deliver efficient and high-speed SNNs going beyond existing implementations. In addition to reducing the energy consumption per spike in the network, SPIKEPro will also develop novel learning strategies and algorithms able to work with reduced number of synaptic connections. These will be possible by exploiting the hardware parameters of the electrical and photonic spiking devices. The outcome of SPIKEPro will have lasting economic, societal and scientific impact. The project will bring ultra fast and efficient neuromorphic hardware into the disparate fields of edge computing, sensor data processing, high-speed control and computational neuroscience.