Feedback Networks

By Amir R. Zamir et al

Published on Aug. 20, 2017

Read the original document by opening this link in a new tab.

Table of Contents

Abstract
1. Introduction
2. Related Work
3. Feedback Networks
3.1. Convolutional LSTM Formulation
3.2. Feedback Module Length
3.3. Temporal Skip Connection
3.4. Taxonomic Prediction
3.5. Episodic Curriculum Learning
3.6. Computation Graph Analysis

Summary

The document discusses Feedback Networks, a learning model in computer vision that utilizes feedback for iterative representation formation. It highlights the advantages of feedback over feedforward approaches, such as enabling early predictions, conforming to hierarchical structures, and supporting Curriculum Learning. The paper presents a general feedback-based learning architecture using existing RNNs and demonstrates its effectiveness. It also explores related works, network formulations, skip connections, taxonomic predictions, and episodic curriculum learning. The computation graph analysis shows the advantages of feedback models in terms of speed and parallelism.