Characteristics of Visual Categorization of Long-Concatenated and Object-Directed Human Actions by a Multiple Spatio-Temporal Scales Recurrent Neural Network Model. (arXiv:1602.01921v1 [cs.CV])

The current paper proposes a novel dynamic neural network model, multiple spatio-temporal scales recurrent neural network (MSTRNN) used for categorization of complex human action pattern in video image. The MSTRNN has been developed by newly introducing recurrent connectivity to a prior-proposed model, multiple spatio-temporal scales neural network (MSTNN) [1] such that the model can learn to extract latent spatio-temporal structures more effectively by developing adequate recurrent contextual dynamics. The MSTRNN was evaluated by conducting a set of simulation experiments on learning to categorize human action visual patterns. The first experiment on categorizing a set of long-concatenated human movement patterns showed that MSTRNN outperforms MSTNN in the capability of learning to extract long-ranged correlation in video image. The second experiment on categorizing a set of object-directed actions showed that the MSTRNN can learn to extract structural relationship between actions and directed-objects. Our analysis on the characteristics of miscategorization in both cases of object-directed action and pantomime actions indicated that the model network developed the categorical memories by organizing relational structure among them. Development of such relational structure is considered to be beneficial for gaining generalization in categorization.

Donate to arXiv

from cs.AI updates on arXiv.org http://ift.tt/1PJbylV
via IFTTT