An Efficient Threshold-Driven Aggregate-Label Learning Algorithm for Multimodal Information Processing

• Published in: IEEE journal of selected topics in signal processing Vol. 14; no. 3; pp. 592 - 602
• Main Authors: Zhang, Malu, Luo, Xiaoling, Chen, Yi, Wu, Jibin, Belatreche, Ammar, Pan, Zihan, Qu, Hong, Li, Haizhou
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.03.2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: aggregate-label learning; Algorithms; Biological neural networks; Biomembranes; Computational modeling; Data processing; Feedback; Machine learning; Membrane potentials; Multilayers; multimodal information; Neural networks; Neurons; Pattern recognition; Signal processing algorithms; Spiking; spiking neural networks; Spiking neurons; synaptic plasticity
• ISSN: 1932-4553; 1941-0484
• DOI: 10.1109/JSTSP.2020.2983547

The aggregate-label learning paradigm tackles the long-standing temporary credit assignment (TCA) problem in neuroscience and machine learning, enabling spiking neural networks to learn multimodal sensory clues with delayed feedback signals. However, the existing aggregate-label learning algorithms only work for single spiking neurons, and with low learning efficiency, which limit their real-world applicability. To address these limitations, in this article, we first propose an efficient threshold-driven plasticity algorithm for spiking neurons, namely ETDP. It enables spiking neurons to generate the desired number of spikes that match the magnitude of delayed feedback signals and to learn useful multimodal sensory clues embedded within spontaneous spiking activities. Furthermore, we extend the ETDP algorithm to support multi-layer spiking neural networks (SNNs), which significantly improves the applicability of aggregate-label learning algorithms. We also validate the multi-layer ETDP learning algorithm in a multimodal computation framework for audio-visual pattern recognition. Experimental results on both synthetic and realistic datasets show significant improvements in the learning efficiency and model capacity over the existing aggregate-label learning algorithms. It, therefore, provides many opportunities for solving real-world multimodal pattern recognition tasks with spiking neural networks.