TY - GEN
T1 - Recurrent Neural Network with Adaptive Gating Timescales Mechanisms for Language and Action Learning
AU - Zhao, Libo
AU - Zhong, Junpei
N1 - Funding Information:
This work is partially supported by PolyU Start-up Grant (ZVUY-P0035417).
Publisher Copyright:
© 2021, Springer Nature Switzerland AG.
PY - 2021/12/2
Y1 - 2021/12/2
N2 - Inspired by the neurons’ differences in membrane time-scales, the multiple timescale recurrent neural network model (MTRNN) adopts the hierarchical architecture with increasing time-scales from bottom to top layers. Based on this idea, the recent adaptive and continuous time recurrent neural networks (ACTRNN) and the gated adaptive continuous time recurrent neural network (GACTRNN) develop the novel learning mechanism on the time-scales. In this paper, we test the performance of GACTRNN using the dataset obtained from a real-world humanoid robot’s object manipulation experiment. By using trainable timescale parameters with the gating mechanism, it can be observed that the GACTRNN can better learn the temporal characteristics of the sequences. Besides, to eliminate the effects of parameters’ overgrowing with a large data-set, we improve the GACTRNN model and propose the MATRNN model. In this model, the sigmoid function is used instead of exponential function. We compare the performances of the CTRNN, GACTRNN and MATRNN models, and find that the GACTRNN and MATRNN models perform better than the CTRNN model with the large-scale dataset. By visualizing the timescales adapting in the training process, we also qualitatively show that the MATRNN model performs better than the GACTRNN model in terms of stability with the dataset.
AB - Inspired by the neurons’ differences in membrane time-scales, the multiple timescale recurrent neural network model (MTRNN) adopts the hierarchical architecture with increasing time-scales from bottom to top layers. Based on this idea, the recent adaptive and continuous time recurrent neural networks (ACTRNN) and the gated adaptive continuous time recurrent neural network (GACTRNN) develop the novel learning mechanism on the time-scales. In this paper, we test the performance of GACTRNN using the dataset obtained from a real-world humanoid robot’s object manipulation experiment. By using trainable timescale parameters with the gating mechanism, it can be observed that the GACTRNN can better learn the temporal characteristics of the sequences. Besides, to eliminate the effects of parameters’ overgrowing with a large data-set, we improve the GACTRNN model and propose the MATRNN model. In this model, the sigmoid function is used instead of exponential function. We compare the performances of the CTRNN, GACTRNN and MATRNN models, and find that the GACTRNN and MATRNN models perform better than the CTRNN model with the large-scale dataset. By visualizing the timescales adapting in the training process, we also qualitatively show that the MATRNN model performs better than the GACTRNN model in terms of stability with the dataset.
KW - Adaptive timescale
KW - Developmental robotics
KW - Membrane time constants
KW - Recurrent neural network
UR - http://www.scopus.com/inward/record.url?scp=85121915975&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-92310-5_47
DO - 10.1007/978-3-030-92310-5_47
M3 - Conference article published in proceeding or book
AN - SCOPUS:85121915975
SN - 9783030923099
T3 - Communications in Computer and Information Science
SP - 405
EP - 413
BT - Neural Information Processing - 28th International Conference, ICONIP 2021, Proceedings
A2 - Mantoro, Teddy
A2 - Lee, Minho
A2 - Ayu, Media Anugerah
A2 - Wong, Kok Wai
A2 - Hidayanto, Achmad Nizar
PB - Springer Science and Business Media Deutschland GmbH
T2 - 28th International Conference on Neural Information Processing, ICONIP 2021
Y2 - 8 December 2021 through 12 December 2021
ER -