Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring the Promise and Limits of Real-Time Recurrent Learning

Kazuki Irie, Anand Gopalakrishnan, Jürgen Schmidhuber

TL;DR

This paper investigates the practical potential of Real-Time Recurrent Learning (RTRL) for real-world sequence processing by enforcing tractability through an element-wise recurrent architecture (eLSTM) and integrating it into an online actor-critic framework (R2AC). It demonstrates that exact RTRL, when restricted to a one-layer eLSTM, can scale to challenging RL benchmarks and outperform certain baselines on memory tasks while using substantially fewer environment frames. The study also candidly discusses limitations, notably the intractability of multi-layer RTRL and dependencies on TBPTT for the vision front-end, outlining directions for hardware optimization and hybrid approaches. Overall, the results show that RTRL can offer practical benefits in realistic tasks, informing future research in architectural design and efficient online gradient computation.

Abstract

Real-time recurrent learning (RTRL) for sequence-processing recurrent neural networks (RNNs) offers certain conceptual advantages over backpropagation through time (BPTT). RTRL requires neither caching past activations nor truncating context, and enables online learning. However, RTRL's time and space complexity make it impractical. To overcome this problem, most recent work on RTRL focuses on approximation theories, while experiments are often limited to diagnostic settings. Here we explore the practical promise of RTRL in more realistic settings. We study actor-critic methods that combine RTRL and policy gradients, and test them in several subsets of DMLab-30, ProcGen, and Atari-2600 environments. On DMLab memory tasks, our system trained on fewer than 1.2 B environmental frames is competitive with or outperforms well-known IMPALA and R2D2 baselines trained on 10 B frames. To scale to such challenging tasks, we focus on certain well-known neural architectures with element-wise recurrence, allowing for tractable RTRL without approximation. Importantly, we also discuss rarely addressed limitations of RTRL in real-world applications, such as its complexity in the multi-layer case.

Exploring the Promise and Limits of Real-Time Recurrent Learning

TL;DR

This paper investigates the practical potential of Real-Time Recurrent Learning (RTRL) for real-world sequence processing by enforcing tractability through an element-wise recurrent architecture (eLSTM) and integrating it into an online actor-critic framework (R2AC). It demonstrates that exact RTRL, when restricted to a one-layer eLSTM, can scale to challenging RL benchmarks and outperform certain baselines on memory tasks while using substantially fewer environment frames. The study also candidly discusses limitations, notably the intractability of multi-layer RTRL and dependencies on TBPTT for the vision front-end, outlining directions for hardware optimization and hybrid approaches. Overall, the results show that RTRL can offer practical benefits in realistic tasks, informing future research in architectural design and efficient online gradient computation.

Abstract

Real-time recurrent learning (RTRL) for sequence-processing recurrent neural networks (RNNs) offers certain conceptual advantages over backpropagation through time (BPTT). RTRL requires neither caching past activations nor truncating context, and enables online learning. However, RTRL's time and space complexity make it impractical. To overcome this problem, most recent work on RTRL focuses on approximation theories, while experiments are often limited to diagnostic settings. Here we explore the practical promise of RTRL in more realistic settings. We study actor-critic methods that combine RTRL and policy gradients, and test them in several subsets of DMLab-30, ProcGen, and Atari-2600 environments. On DMLab memory tasks, our system trained on fewer than 1.2 B environmental frames is competitive with or outperforms well-known IMPALA and R2D2 baselines trained on 10 B frames. To scale to such challenging tasks, we focus on certain well-known neural architectures with element-wise recurrence, allowing for tractable RTRL without approximation. Importantly, we also discuss rarely addressed limitations of RTRL in real-world applications, such as its complexity in the multi-layer case.
Paper Structure (42 sections, 52 equations, 2 figures, 5 tables)

This paper contains 42 sections, 52 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Preliminaries.
  4. Real-Time Recurrent Learning (RTRL).
  5. Method
  6. RTRL for LSTM with Element-wise Recurrence (eLSTM)
  7. Real-Time Recurrent Actor-Critic Policy Gradient Algorithm (R2AC)
  8. Experiments
  9. DMLab Memory Tasks.
  10. ProcGen.
  11. Atari.
  12. Limitations and Discussion
  13. Conclusion
  14. Derivations
  15. RTRL for LSTM with Element-wise Recurrence
  16. ...and 27 more sections

Figures (2)

  • Figure 1: Training curves on DMLab-30rooms_select_nonmatching_object (Non-matching) and rooms_watermaze (Watermaze), and ProcgenChaser environments.
  • Figure 2: Learning curves on five Atari environments