Timezone: »

Reviving and Improving Recurrent Back-Propagation
Renjie Liao · Yuwen Xiong · Ethan Fetaya · Lisa Zhang · Kijung Yoon · Zachary S Pitkow · Raquel Urtasun · Richard Zemel

Thu Jul 12 09:15 AM -- 12:00 PM (PDT) @ Hall B #178

In this paper, we revisit the recurrent back-propagation (RBP) algorithm, discuss the conditions under which it applies as well as how to satisfy them in deep neural networks. We show that RBP can be unstable and propose two variants based on conjugate gradient on the normal equations (CG-RBP) and Neumann series (Neumann-RBP). We further investigate the relationship between Neumann-RBP and back propagation through time (BPTT) and its truncated version (TBPTT). Our Neumann-RBP has the same time complexity as TBPTT but only requires constant memory, whereas TBPTT's memory cost scales linearly with the number of truncation steps. We examine all RBP variants along with BPTT and TBPTT in three different application domains: associative memory with continuous Hopfield networks, document classification in citation networks using graph neural networks and hyperparameter optimization for fully connected networks. All experiments demonstrate that RBPs, especially the Neumann-RBP variant, are efficient and effective for optimizing convergent recurrent neural networks.

Author Information

Renjie Liao (University of Toronto)
Yuwen Xiong (Uber ATG / University of Toronto)
Ethan Fetaya (University of Toronto)
Lisa Zhang (University of Toronto)
Kijung Yoon (Rice University)
Zachary S Pitkow (Baylor College of Medicine / Rice University)
Raquel Urtasun (University of Toronto)
Richard Zemel (Vector Institute)

Related Events (a corresponding poster, oral, or spotlight)

More from the Same Authors