多模态融合的听觉注意基序脉冲神经网络模型代码

网络架构和学习原则在开发人工神经网络（ANNs）的复杂认知能力方面至关重要。脉冲神经网络（SNNs）是ANNs的一个子集，其中纳入了额外的生物学特性，例如动态脉冲神经元、生物学指定的架构以及高效和有用的范例。在这里，我们更关注SNNs中的网络架构，例如从生物网络中借用的称为3节点网络基序的元运算符。我们提出了一个改进的脉冲神经网络（M-SNN），该网络进一步证实了其在解释关键认知现象方面的效率，例如鸡尾酒会效应（典型的抗噪声语音识别任务）和麦高（McGurk）效应（典型的跨感官整合任务）。对于M-SNN，基序拓扑结构是通过整合空间和时间基序获得的。这些空间和时间基序首先从空间（例如MNIST）和时间（例如TIDigits）数据集的预训练中生成，然后应用于之前介绍的两个认知效应任务。实验结果表明，这种基序拓扑结构具有更低的计算成本、更高的准确性和更好地解释这两个效应的一些关键现象，例如新概念生成和抗背景噪声。这种介观网络基序拓扑结构在未来有很大的发展空间。

Network architectures and learning principles are crucial for developing the sophisticated cognitive capabilities of artificial neural networks (ANNs). Spiking neural networks (SNNs) are a subset of ANNs that incorporate additional biological properties, such as dynamic spiking neurons, biologically specified architectures, and efficient and practical paradigms. Herein, we focus more on the network architectures within SNNs, such as the meta-operators called 3-node network motifs borrowed from biological networks. We propose a modified spiking neural network (M-SNN), which further validates its efficiency in explaining key cognitive phenomena, such as the cocktail party effect (a typical noise-resistant speech recognition task) and the McGurk effect (a typical cross-sensory integration task). For the M-SNN, the motif topology is obtained by integrating spatial and temporal motifs. These spatial and temporal motifs are first generated via pre-training on spatial (e.g., MNIST) and temporal (e.g., TIDigits) datasets, and then applied to the two cognitive effect tasks introduced previously. Experimental results demonstrate that this motif topology features lower computational cost, higher accuracy, and better ability to explain some key phenomena of these two effects, such as novel concept generation and background noise resistance. This mesoscale network motif topology has great potential for future development.

搜集汇总

背景与挑战

背景概述

该数据集提供了一个多模态听觉注意的基序拓扑脉冲神经网络（M-SNN）模型代码，由中国科学院自动化研究所于2023年发布。它通过整合空间和时间基序，旨在解释鸡尾酒会效应和McGurk效应等认知现象，具有较低计算成本和较高准确性的特点，适用于脑科学和人工智能领域的研究。