Attention-Based Deep Spiking Neural Networks for Temporal Credit Assignment Problems.

Area of Science:

• Machine Learning
• Computational Neuroscience
• Signal Processing

Background:

• Temporal credit assignment (TCA) is crucial for identifying predictive features amidst background noise in biological and machine learning systems.
• Aggregate-label (AL) learning addresses TCA by linking neural spikes with delayed feedback, but current methods lack multi-timestep analysis and quantitative evaluation.
• Existing AL algorithms' single-timestep focus limits their applicability in real-world scenarios requiring analysis of complex temporal dependencies.

Purpose of the Study:

• To propose a novel attention-based temporal credit assignment (ATCA) algorithm for enhanced feature detection.
• To introduce a minimum editing distance (MED)-based quantitative evaluation method for TCA problems.
• To overcome the limitations of existing AL learning algorithms by incorporating multi-timestep information and robust evaluation.

Main Methods:

• Developed an ATCA algorithm utilizing an attention mechanism within its loss function to process information from spike clusters.
• Implemented a minimum editing distance (MED) metric to quantitatively assess the similarity between spike trains and target clue flows.
• Validated the ATCA algorithm and MED evaluation on diverse datasets including MedleyDB, TIDIGITS, and DVS128-Gesture.

Main Results:

• The proposed ATCA algorithm demonstrated superior performance in temporal credit assignment tasks.
• The MED-based evaluation method provided a quantitative measure for TCA problem assessment.
• Experimental results confirmed that ATCA achieves state-of-the-art (SOTA) performance compared to existing AL learning algorithms across multiple recognition tasks.

Conclusions:

• The ATCA algorithm effectively addresses the limitations of current AL learning methods by considering multi-timestep information.
• The MED metric offers a reliable quantitative evaluation for TCA problems.
• The study establishes a new benchmark for TCA performance in machine learning applications.