Event-Based Feature Extraction Using Adaptive Selection Thresholds.

Area of Science:

• Machine Learning
• Neuromorphic Engineering
• Computer Science

Background:

• Unsupervised feature extraction is crucial for machine learning.
• Adapting these algorithms for event-based neuromorphic hardware often requires performance-compromising simplifications.
• Existing methods lack useful intermediary signals tailored for hardware constraints.

Purpose of the Study:

• To propose a novel event-based feature extraction method addressing hardware limitations.
• To enable simpler implementation of network homeostasis and provide weight convergence signals.
• To develop a heuristic for network size selection and predict classification accuracy.

Main Methods:

• Developed an algorithm with adaptive selection thresholds to manage information loss.
• Utilized selection threshold behavior and network output to indicate weight convergence.
• Introduced a heuristic for network size selection based on noise events.
• Evaluated network activation patterns for predicting classification accuracy.

Main Results:

• The proposed method allows simpler network homeostasis with minimal information loss.
• Network behavior signals weight convergence without direct weight access.
• A novel heuristic aids in network size selection.
• Predicted classification accuracy using network activation patterns, enabling rapid optimization.
• Demonstrated performance gains on N-MNIST and airplane datasets.

Conclusions:

• The novel event-based feature extraction method enhances neuromorphic system efficiency and optimization.
• Adaptive thresholds offer a practical trade-off for hardware implementation.
• The method provides valuable insights into network dynamics and performance prediction.