Efficient Online Learning with Predictive Coding Networks: Exploiting Temporal Correlations

Darius Masoum Zadeh-Jousdani, Elvin Hajizada, Eyke Hüllermeier

Introduction

Deep learning has traditionally relied on backpropagation (BP) to train neural networks. While highly effective, BP is often criticized for its biological implausibility, as it requires global information to update synaptic weights. Predictive Coding (PC) offers a compelling alternative, utilizing local "Hebbian" updates that are more aligned with biological processes and ideal for specialized neuromorphic hardware.

However, standard Predictive Coding faces a significant "speed limit." For every input, the network must undergo multiple internal inference iterations to minimize prediction errors before weight updates can occur. In real-time robotic applications processing video streams, re-calculating these internal states from scratch for every frame is computationally expensive and redundant, given the high temporal correlation between consecutive frames.

The PCN-TA Framework: Temporal Amortization

To address these inefficiencies, researchers introduced PCN-TA (Predictive Coding Network with Temporal Amortization). This approach leverages the temporal continuity of streaming data to bypass redundant computations.

In a standard Predictive Coding Network (PCN), hidden states are reset for each new input. PCN-TA, conversely, "remembers" internal neural states across frames. When a new frame arrives, the network performs a standard feedforward pass but then reinstates the hidden states saved from the previous frame—a "warm start." It then performs iterative updates to minimize Variational Free Energy (VFE), aligning top-down predictions with bottom-up sensory input. Because consecutive frames are similar, the previous state provides a high-quality initial guess, allowing the network to reach equilibrium significantly faster.

Experimental Evaluation

The effectiveness of PCN-TA was evaluated using the COIL-20 robotic perception dataset. COIL-20 contains sequential video frames of 20 objects being rotated, providing the smooth temporal transitions necessary to test online learning and temporal amortization.

Performance and Accuracy

The study compared PCN-TA against standard PCNs and traditional backpropagation networks (BPN). As shown in Figure 1, PCN-TA with 100 inference steps not only tracks the accuracy of backpropagation but eventually exceeds it in later training epochs. Notably, PCN-TA with only 50 inference iterations outperformed a standard PC network using 100 iterations, demonstrating the power of the "warm start" mechanism.

Figure 1: Test accuracy comparison between BPN, standard PCN, and PCN-TA across training epochs.

Computational Efficiency and Sparsity

A key advantage of PCN-TA is its computational efficiency. By starting with a better estimate of hidden states, the resulting prediction error is smaller. This leads to weight update sparsity, where the network only needs to adjust synaptic weights for the specific parts of the input that have changed.

As illustrated in Figure 2, PCN-TA requires significantly fewer weight updates per frame compared to both standard PC and traditional backpropagation. Specifically, PCN-TA achieved a 10% reduction in weight updates compared to BP and a 50% reduction in inference steps compared to baseline PC networks.

Figure 2: Average number of weight updates per frame for Backpropagation, standard PC, and PCN-TA.

Discussion and Key Insights

The primary innovation of PCN-TA is treating temporal continuity as a computational shortcut. By exploiting the fact that the world changes incrementally, the model avoids redundant "re-thinking" of its internal state. This leads to several key insights:

  • Efficiency through Continuity: Temporal amortization allows for high accuracy with fewer computational steps.
  • Robust Representations: The iterative inference process in PCN-TA may provide a more robust representation for online learning than the single-pass nature of standard BP.
  • Hardware Compatibility: The local update rules and reduced computational load make PCN-TA particularly well-suited for deployment on edge robotics and neuromorphic hardware.

Limitations and Future Directions

While promising, the current study has limitations. The COIL-20 dataset is relatively small and lacks the complexity of real-world environments. Furthermore, the model employs a "fixed-prediction assumption" to approximate backpropagation, which deviates slightly from pure predictive coding theory.

Future research will focus on testing PCN-TA with high-resolution, complex video streams. The ultimate goal is to deploy the algorithm on neuromorphic hardware to quantify actual power savings in real-time robotic adaptation scenarios.

Summary updated

BibTeX
@inproceedings{zadehjousdani2025efficient,
  author = {Darius Masoum Zadeh-Jousdani and Elvin Hajizada and Eyke Hüllermeier},
  title = {{Efficient Online Learning with Predictive Coding Networks: Exploiting Temporal Correlations}},
  booktitle = {EdgeAI4R @ IROS},
  year = {2025},
  url = {https://openreview.net/forum?id=GKfyGppO1D},
}
Figure