Attention and Recurrence for DAgger using Pytorch
MEng. Project Electrical and Computer Engineering University of Toronto, St. George Ekeledirichukwu C. Nnorom
In this project the DAgger algorithm is augmented with Recurrent Convolutional Neural Networks and Attention Networks for the purpose of imitation learning. A prominent difference is between the approaches here is most CNN are typically a feed-forward architecture – an incomplete representation of the human visual system which has ubiquitous recurrent connections. Despite static input, the activities of an RCNN unit evolve over time so that the activity of each unit is modulated by the activities of its neighboring units. This enhances the model’s ability to integrate the context information, which is important for object recognition. The models trained with Attention Gates (AG) implicitly learn to suppress irrelevant regions in an input image while highlighting useful features. The result of these approaches is a more robust policy for Imitation Learning. Keywords: Recurrence, Attention Gates, Dagger