Neuromorphic Text Emotion Recognition: Harnessing Bio- Inspired Computing for Energy-Efficient and Robust Solution

Due to its strong biological similarity and low energy consumption, neuromorphic computing has become a popular new computing architecture. Different from Artificial Neural Networks (ANNs), Spiking Neural Networks (SNNs) are the main information processing paradigm in this novel architecture, which is an exceptionally sparse way of information transfer. However, there is no free lunch in the world, and sparsity also leads to the phenomenon of the dead neuron in the existing model output (the output layer cannot emit spikes), which limits the effective application of SNNs. To address this issue, this paper takes text emotion recognition as a starting point and employs SNN s for modeling. We have explored two ways to obtain SNN models, the ANN conversed SNN method or trained SNN by surrogate gradient. Population strategy is commonly utilized in trained SNN by the surrogate gradient. Through experimental analysis of the model, it was found that the structure can be borrowed to alleviate the problem of too many dead neurons in the ANN conversed SNN method. Experimental results demonstrate that introducing the population strategy into the ANN conversed SNN structure can not only improve the accuracy of the model but also enhance its robustness to low noise. Using Gaussian noise with a variance of 0.1 to interfere with CMU-MOSI testing data, the proposed SNN model in this study experiences a mere 5% decline in accuracy, as opposed to the substantial 47% decline observed in ANNs.