Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Efficient Online Learning with Predictive Coding Networks: Exploiting Temporal Correlations

Darius Masoum Zadeh-Jousdani, Elvin Hajizada, Eyke Hüllermeier

TL;DR

The paper addresses online edge learning for robotics by proposing Predictive Coding Network with Temporal Amortization (PCN-TA), a biologically plausible alternative to backpropagation that exploits temporal correlations to reduce computation. PCN-TA preserves latent states across sequential frames, lowering inference iterations and weight updates while maintaining competitive accuracy on the COIL-20 dataset. Compared to backpropagation, PCN-TA achieves about 10% fewer weight updates and uses roughly 50% fewer inference steps than a baseline PCN, with accuracy approaching that of standard methods in later epochs. This approach offers a viable path toward efficient, real-time, edge deployment and potential neuromorphic hardware implementations for continual learning in robotics.

Abstract

Robotic systems operating at the edge require efficient online learning algorithms that can continuously adapt to changing environments while processing streaming sensory data. Traditional backpropagation, while effective, conflicts with biological plausibility principles and may be suboptimal for continuous adaptation scenarios. The Predictive Coding (PC) framework offers a biologically plausible alternative with local, Hebbian-like update rules, making it suitable for neuromorphic hardware implementation. However, PC's main limitation is its computational overhead due to multiple inference iterations during training. We present Predictive Coding Network with Temporal Amortization (PCN-TA), which preserves latent states across temporal frames. By leveraging temporal correlations, PCN-TA significantly reduces computational demands while maintaining learning performance. Our experiments on the COIL-20 robotic perception dataset demonstrate that PCN-TA achieves 10% fewer weight updates compared to backpropagation and requires 50% fewer inference steps than baseline PC networks. These efficiency gains directly translate to reduced computational overhead for moving another step toward edge deployment and real-time adaptation support in resource-constrained robotic systems. The biologically-inspired nature of our approach also makes it a promising candidate for future neuromorphic hardware implementations, enabling efficient online learning at the edge.

Efficient Online Learning with Predictive Coding Networks: Exploiting Temporal Correlations

TL;DR

The paper addresses online edge learning for robotics by proposing Predictive Coding Network with Temporal Amortization (PCN-TA), a biologically plausible alternative to backpropagation that exploits temporal correlations to reduce computation. PCN-TA preserves latent states across sequential frames, lowering inference iterations and weight updates while maintaining competitive accuracy on the COIL-20 dataset. Compared to backpropagation, PCN-TA achieves about 10% fewer weight updates and uses roughly 50% fewer inference steps than a baseline PCN, with accuracy approaching that of standard methods in later epochs. This approach offers a viable path toward efficient, real-time, edge deployment and potential neuromorphic hardware implementations for continual learning in robotics.

Abstract

Robotic systems operating at the edge require efficient online learning algorithms that can continuously adapt to changing environments while processing streaming sensory data. Traditional backpropagation, while effective, conflicts with biological plausibility principles and may be suboptimal for continuous adaptation scenarios. The Predictive Coding (PC) framework offers a biologically plausible alternative with local, Hebbian-like update rules, making it suitable for neuromorphic hardware implementation. However, PC's main limitation is its computational overhead due to multiple inference iterations during training. We present Predictive Coding Network with Temporal Amortization (PCN-TA), which preserves latent states across temporal frames. By leveraging temporal correlations, PCN-TA significantly reduces computational demands while maintaining learning performance. Our experiments on the COIL-20 robotic perception dataset demonstrate that PCN-TA achieves 10% fewer weight updates compared to backpropagation and requires 50% fewer inference steps than baseline PC networks. These efficiency gains directly translate to reduced computational overhead for moving another step toward edge deployment and real-time adaptation support in resource-constrained robotic systems. The biologically-inspired nature of our approach also makes it a promising candidate for future neuromorphic hardware implementations, enabling efficient online learning at the edge.

Paper Structure

This paper contains 8 sections, 2 figures, 2 algorithms.

Table of Contents

  1. INTRODUCTION
  2. Mathematical Formulations of PCN-TA
  3. Methodology
  4. Experiments
  5. Accuracy results
  6. Weight Update Sparsity
  7. Discussion
  8. Conclusion

Figures (2)

  • Figure 1: The results of the image classification task show that backpropagation (red line) consistently achieves the highest average accuracy across all epochs. In contrast, the standard PCN with 100 inference iterations (blue line) performs the worst in nearly every epoch. Notably, the PCN-TA with only 50 inference iterations (brown line) outperforms the standard PCN, despite requiring fewer iterations.
  • Figure 2: This chart presents the average number of weight updates per frame across all epochs. The y-axis shows the average updates per frame, measured in millions. Backpropagation begins with nearly 1.7 million updates in the first epoch, with the PCN exhibiting a similar value. In contrast, the PCN-TA starts with a considerably lower number of updates and maintains this advantage throughout training. Overall, the PCN and backpropagation display comparable update counts, whereas the PCN-TA consistently achieves the lowest number across all epochs.