Conceptor-Augmented Regularization for RNN and LSTM Networks

Recurrent Neural Network (RNN) and its variants, such as Long Short-Term Memory (LSTM), have achieved remarkable success in sequential data processing tasks. The representations learned by their hidden layers often require further regularization to enhance model performance and robustness. In this paper, we propose a novel Conceptor-based regularization mechanism for improving the hidden state representations of RNN and LSTM networks. Unlike previous works that primarily treat Conceptors as classifiers, we introduce a dynamic Conceptor layer integrated within the hidden states of RNNs and LSTM networks, termed Conceptor-Augmented RNN (CA-RNN) and Conceptor-Augmented LSTM (CA-LSTM). We evaluate the proposed method on the task of classifying speakers using Japanese vowel recordings, comparing standard RNN and LSTM models with those augmented by our Conceptor-based regularization. Our results demonstrate that incorporating the proposed dynamic Conceptor mechanism significantly improves classification performance and noise resistance, highlighting its effectiveness as a hidden-state regularizer for sequential models.