Deep active object recognition by joint label and action prediction

Mohsen Malmir, Karan Sikka, Deborah Forster, Ian R. Fasel, Javier R. Movellan, Garrison W. Cottrell

Abstract: Highlights•A deep convolutional neural network is used for active object recognition.•The network outputs object label and the action to be performed on the object.•Cost function includes reinforcement learning and label prediction costs.•A generative Dirichlet model is embedded in the network for state encoding.•Training is done by tuning the visual features for label and action prediction.