Learning off-line vs. on-line models of interactive multimodal behaviors with recurrent neural networks

Abstract

Human interactions are driven by multi-level perception-action loops. Interactive behavioral models are typically built using rule-based methods or statistical approaches such as Hidden Markov Model (HMM), Dynamic Bayesian Network (DBN), etc. In this paper, we present the multimodal interactive data and our behavioral model based on recurrent neural networks, namely Long-Short Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) models. Speech, gaze and gestures of two subjects involved in a collaborative task are here jointly modeled. The results show that the proposed deep neural networks are more effective than the conventional statistical methods in generating appropriate overt actions for both on-line and off-line prediction tasks.