基于时空信息融合的多任务脉冲神经网络研究

How to get the efficient representation and integration of temporal and spatial information is one of the key research points in general artificial intelligence area. In biological plausible spiking neural network, the temporal information processing (e.g. biological integrated-and-fire neuron and synaptic delay) and spatial information processing (e.g. membrane potential coding, fire rate coding and information abstraction from the level of networks) could be combined together for integrated information representation. Based on this characteristic, we will construct a spiking neural network which could handle different kinds of tasks (i.e. classification task which focuses on spatial information and prediction task which focuses on temporal information). On the perspective of structure, we will get the inspirations from structures of Hippocampus; on the perspective of functions, very different from current artificial back propagation tuning method, we will tune the spiking neural network directly by multi-scale biological plasticity rules, such as Hebb's law, spike-timing dependent plasticity (STDP), short-term plasticity (STP), biological self-organized back propagation. We will further discuss effectiveness of neuron types and neuron numbers to the performance of spiking neural network. Finally, based on the constructed Baxter (with two arms) robot environment, the proposed spatial-temporal integrated spiking model will be verified on two tasks at the same time (i.e. visual object learning task and motion sequence learning task).

如何对时间和空间信息进行有效表征和融合是目前通用人工智能系统重点关注的问题之一。在高仿生的生物脉冲神经网络中，神经元的积分放电、突触的传递延迟等时间信息运算，可以和神经元的膜电位编码、放电率编码、网络层面信息抽象等空间信息运算结合起来，形成一种综合信息表达。本课题将以此为基础，建立可以同时处理两大类认知任务（侧重于空间信息的分类任务和侧重于时序信息的预测任务）的通用脉冲神经网络模型。在结构方面，借鉴海马区不同亚区的网络连接模式；在功能方面，与传统的采用误差反传的优化方法不同，我们将采用类生物的多尺度神经可塑性机制来训练脉冲神经网络，如Hebb法则、时序依赖的突触可塑性、短时突触可塑性、生物自组织误差反馈等。进一步的，我们将探讨不同神经元类型及比例等因素对脉冲网络优化的影响。最后，在Baxter双臂机器人上验证时空融合模型同时在视觉目标学习、动作序列学习等不同任务上的有效性。

如何对时间和空间信息进行有效表征和融合是目前通用人工智能系统重点关注的问题之一。本项目基于高仿生的生物脉冲神经网络，研究微观尺度神经元的动态积分放电，介观尺度的网络环路及可塑性传播，宏观尺度的时空信息融合等。项目建立了可以同时处理两大类认知任务的脉冲神经网络模型，包括空间信息分类任务和时序信息预测任务。实验结果验证了系列脉冲神经网络模型的时空信息处理优势，未来有望在低能耗、小型化、类人机器人等神经形态设备应用上发挥优势。我们在在信息高效编码、网络构建、节点设置、学习方法优化等多方面达到了预期进展，相关成果发表在Science子刊Science Advances、IEEE会刊TNNLS、Frontiers in Neuroscience等期刊，以及AAAI、ICASSP等会议上。