Abstract & Overview
Spike-driven neural networks, inspired by the unique property of brain event-driven processing, lead to significant advantages in terms of sparse data representation, efficient parallel processing, and low power consumption. These characteristics are specifically beneficial for dealing with large volumes of text data, particularly in sentiment analysis and text classification, where the capability to process multiple input data simultaneously can lead to the more accurate and faster results. Moreover the use of memristor crossbar arrays further enhances the system's parallelism and efficiency by providing non-volatile, high density integration of synaptic weights, equipping scalable implementations suitable for edge computing and its applications. The proposed approach leverages the robustness and adaptability of spike-driven neural networks in learning features from sequential and temporal data, which is vital for natural language processing (NLP). By employing multi-head attention (MHA) mechanism and spiking-ANN, the model can attend to different perspectives of the input text data concurrently, capturing both global and local dependencies. This capability is especially beneficial in sentiment analysis and text classification, where the expressions of sentiments can change widely across different contexts.