Article ID | Journal | Published Year | Pages | File Type |
---|---|---|---|---|
403848 | Neural Networks | 2015 | 15 Pages |
The multiple timescale recurrent neural network (MTRNN) model is a useful tool for recording and regenerating a continuous signal for dynamic tasks. However, our research shows that the MTRNN model is difficult to use for the classification of multiple types of motion when observing a human action. Therefore, in this paper, we propose a new supervised MTRNN model for handling the issue of action classification. Instead of setting the initial states, we define a group of slow context nodes as “classification nodes.” The supervised MTRNN model provides both prediction and classification outputs simultaneously during testing. Our experiment results show that the supervised MTRNN model inherits the basic function of an MTRNN and can be used to generate action signals. In addition, the results show that the robustness of the supervised MTRNN model is better than that of the MTRNN model when generating both action sequences and action classification tasks.