Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Effect of Optimizer, Initializer, and Architecture of Hypernetworks on Continual Learning from Demonstration

Sayantan Auddy, Sebastian Bergner, Justus Piater

TL;DR

The paper addresses how optimizer, initializer, and architecture choices affect hypernetwork-based continual learning from demonstration (CLfD) for real-world robotic trajectories. It employs hypernetworks and chunked hypernetworks to generate target networks, evaluating NODE and $s$NODE targets across the RoboTasks9 LfD benchmark. Key findings show adaptive optimizers ($Adam$ and $RMSProp$) outperform SGD, hypernetwork-specific initializers offer no clear benefit over a strong default like $Kaiming$, and $s$NODE targets yield largely architecture-insensitive performance, highlighting robustness in real-world continual learning. The work provides practical guidance for designing hypernetwork-based CLfD systems and releases open-source code for reproducibility.

Abstract

In continual learning from demonstration (CLfD), a robot learns a sequence of real-world motion skills continually from human demonstrations. Recently, hypernetworks have been successful in solving this problem. In this paper, we perform an exploratory study of the effects of different optimizers, initializers, and network architectures on the continual learning performance of hypernetworks for CLfD. Our results show that adaptive learning rate optimizers work well, but initializers specially designed for hypernetworks offer no advantages for CLfD. We also show that hypernetworks that are capable of stable trajectory predictions are robust to different network architectures. Our open-source code is available at https://github.com/sebastianbergner/ExploringCLFD.

Effect of Optimizer, Initializer, and Architecture of Hypernetworks on Continual Learning from Demonstration

TL;DR

The paper addresses how optimizer, initializer, and architecture choices affect hypernetwork-based continual learning from demonstration (CLfD) for real-world robotic trajectories. It employs hypernetworks and chunked hypernetworks to generate target networks, evaluating NODE and NODE targets across the RoboTasks9 LfD benchmark. Key findings show adaptive optimizers ( and ) outperform SGD, hypernetwork-specific initializers offer no clear benefit over a strong default like , and NODE targets yield largely architecture-insensitive performance, highlighting robustness in real-world continual learning. The work provides practical guidance for designing hypernetwork-based CLfD systems and releases open-source code for reproducibility.

Abstract

In continual learning from demonstration (CLfD), a robot learns a sequence of real-world motion skills continually from human demonstrations. Recently, hypernetworks have been successful in solving this problem. In this paper, we perform an exploratory study of the effects of different optimizers, initializers, and network architectures on the continual learning performance of hypernetworks for CLfD. Our results show that adaptive learning rate optimizers work well, but initializers specially designed for hypernetworks offer no advantages for CLfD. We also show that hypernetworks that are capable of stable trajectory predictions are robust to different network architectures. Our open-source code is available at https://github.com/sebastianbergner/ExploringCLFD.
Paper Structure (4 sections, 2 figures)

This paper contains 4 sections, 2 figures.

Table of Contents

  1. Introduction
  2. Background
  3. Experiments and Results
  4. Conclusion

Figures (2)

  • Figure 1: DTW errors (lower is better) of different optimizers (top), and initializers (bottom) for 5 independent runs. For reference, the dotted brown line shows the best possible median DTW score from auddy_snode (when each task is learned with a separate model).
  • Figure 2: Effect of hypernetwork depth (y-axis) and target network depth (x-axis) on continual learning from demonstration. Each heatmap corresponds to a different hypernetwork type. Circled numbers show the best DTW for each hypernetwork. Colors are scaled logarithmically. Median values over 5 independent runs are shown.