Dynamic Attention Networks for Task Oriented Grounding

TL;DR

This paper introduces a Dynamic Attention Network for task-oriented grounding in visual environments, improving multi-modal fusion of text and visual data for robotic control without prior domain knowledge.

Contribution

The work presents a novel end-to-end trainable architecture that models dynamic attention in continuous 3D visual worlds, enhancing grounding and policy learning for robots.

Findings

Dynamic Attention improves grounding accuracy.

Using 1D convolution accelerates network convergence.

LSTM cell-state effectively models dynamic attention.

Abstract

In order to successfully perform tasks specified by natural language instructions, an artificial agent operating in a visual world needs to map words, concepts, and actions from the instruction to visual elements in its environment. This association is termed as Task-Oriented Grounding. In this work, we propose a novel Dynamic Attention Network architecture for the efficient multi-modal fusion of text and visual representations which can generate a robust definition of state for the policy learner. Our model assumes no prior knowledge from visual and textual domains and is an end to end trainable. For a 3D visual world where the observation changes continuously, the attention on the visual elements tends to be highly co-related from a one-time step to the next. We term this as "Dynamic Attention". In this work, we show that Dynamic Attention helps in achieving grounding and also aids in the policy learning objective. Since most practical robotic applications take place in the real world where the observation space is continuous, our framework can be used as a generalized multi-modal fusion unit for robotic control through natural language. We show the effectiveness of using 1D convolution over Gated Attention Hadamard product on the rate of convergence of the network. We demonstrate that the cell-state of a Long Short Term Memory (LSTM) is a natural choice for modeling Dynamic Attention and shows through visualization that the generated attention is very close to how humans tend to focus on the environment.