Quantifying Temporal Entropy in Neuromorphic Memory Forgetting: Exploring Advanced Forgetting Models for Robust Long-Term Information Storage

Shane Harrigan, Sonya Coleman, Dermot Kerr, Justin Quinn, Kyle Madden, Shuo Liu, Leeanne LindsaySSCI 2023 · DOI

Abstract

Neuromorphic sensors generate high-frequency, asynchronous event streams that require efficient memory management to maintain real-time processing capabilities. Traditional memory management in data structures like the Reduction-Over-Time (ROT) tree relies on simple exponential decay, which often fails to distinguish critical information from noise. This paper evaluates four biologically inspired forgetting models—Adaptive, Emotional, Context-Dependent, and Multi-dimensional—integrated into a corner detection task. Our findings demonstrate that the Multi-dimensional model, which incorporates relevance, significance, and spatial cues, achieves a superior F1 score of 94.8%. This research suggests that multi-factor forgetting mechanisms are essential for robust long-term information storage in neuromorphic systems.

Introduction

Neuromorphic sensors, such as event-based cameras, mimic biological vision by transmitting data only when individual pixels detect changes in light intensity. This approach offers extreme efficiency but produces a massive stream of asynchronous events. To manage this data, researchers utilize the Reduction-Over-Time (ROT) tree as a form of short-term memory.

Historically, the ROT tree has employed a "forgetting" mechanism based on thermodynamic cooling or simple exponential decay: older data is deleted regardless of its informational value. However, biological memory suggests that importance is not solely a function of time. This study explores "smarter" forgetting models that incorporate nuances like information difficulty and environmental context to improve the retention of significant visual features.

Methodology

The study introduces four novel forgetting models compared against the baseline ROT model. These models were integrated into the ROT-Harris corner detection algorithm to determine which events should be retained to define object boundaries.

  1. Adaptive Forgetting: Introduces a difficulty factor ($D$) to adjust retention based on the complexity of processing specific information.
  2. Emotional Memory Enhancement: Uses an emotional valence factor ($E$) as a probabilistic proxy for high-intensity events that warrant vivid retention.
  3. Context-Dependent Memory: Employs a context factor ($P$) based on surrounding activity, mimicking environmental recall triggers.
  4. Multi-dimensional Memory Forgetting: The most sophisticated model, combining relevance ($r$), significance ($s$), and spatial cues ($c$) to dynamically calculate decay rates.
Table 1: Parameter values and variables assigned to the proposed forgetting functions.

Experimental Setup

The models were tested using the shapes dataset from the DAVIS240-C neuromorphic database. This dataset provides event-based vision data at 240 × 180 resolution with ground truth frames, enabling precise calculation of accuracy and F1 scores. The primary objective was to evaluate how effectively each model could maintain corner detection accuracy while discarding irrelevant event noise.

Results and Discussion

The experimental results highlight a significant performance gap between simple decay and multi-factor models.

  • The Multi-dimensional Model: This model was the top performer, achieving an F1 score of 94.8% and an accuracy of 90.1%. By utilizing spatial cues and relevance, it maintained a highly accurate "mental map" of the scene with a decision speed of 102 nanoseconds.
  • Adaptive Forgetting: While this model was the fastest (35ns decision time), its accuracy was poor (29.8%), suggesting that difficulty alone is insufficient for robust memory management.
  • Context-Dependent Failure: This model performed the worst (19.9% accuracy), likely because "context" is difficult to define mathematically within a raw, high-speed event stream, leading to inconsistent data retention.
Figure 1: Visual comparison of corner detection performance across the baseline ROT and the four proposed forgetting models.

Conclusion

This research confirms that biological realism in data management—specifically the transition from time-based decay to multi-dimensional relevance—significantly enhances the performance of neuromorphic memory systems. While the Multi-dimensional model requires more complex parameter tuning, its ability to distinguish signal from noise makes it the most viable candidate for long-term information storage in autonomous systems. Future work will focus on automating parameter tuning and testing these models in real-world robotic navigation tasks.

Summary updated

BibTeX
@inproceedings{harrigan2023quantifying,
  author = {Shane Harrigan and Sonya Coleman and Dermot Kerr and Justin Quinn and Kyle Madden and Shuo Liu and Leeanne Lindsay},
  title = {{Quantifying Temporal Entropy in Neuromorphic Memory Forgetting: Exploring Advanced Forgetting Models for Robust Long-Term Information Storage}},
  booktitle = {SSCI},
  year = {2023},
  doi = {10.1109/ssci52147.2023.10372015},
}
Figure