Visual Place Recognition Using Rate-Encoded Spiking Neural Networks with Discrete STDP Learning

This paper addresses the performance gap in Spiking Neural Network (SNN) based Visual Place Recognition (VPR), where high classification accuracy often fails to translate into the high retrieval precision required for reliable autonomous navigation. The authors move away from continuous-time ODE-based simulations—which can introduce temporal leakage and ambiguity—to a discrete, tensor-native implementation using PyTorch and snnTorch. This framework allows for a controlled, factorial ablation of three key design decisions: the neuron assignment strategy, the isolation of neuronal states between queries, and the use of temporal sequence aggregation.

• Assignment Strategy: The authors introduce a deterministic, closed-form tensor pipeline for neuron assignment. By moving from a standard argmax approach to a probabilistic assignment that accounts for response magnitude and neuron selectivity, they significantly improve R@100P without sacrificing overall classification accuracy.
• State Isolation: By explicitly resetting membrane potentials and synaptic traces to zero at the start of every query, the authors eliminate temporal spillover, which provides a measurable, consistent boost to retrieval performance.
• Sequence Aggregation: The study demonstrates that aggregating probability outputs over a sliding window of $k$ consecutive frames, compensated for robot velocity, is a powerful technique. For constant-velocity traversals, this method reaches 100% R@100P at $k=5$ with minimal additional latency (0.20 ms), proving that sequence-level context is sufficient to overcome single-frame perceptual aliasing.

Reliable loop closure is a cornerstone of autonomous SLAM systems. By formalizing these SNN design choices into a GPU-native, deterministic pipeline, the authors provide a clear path for deploying high-precision, unsupervised VPR on neuromorphic hardware. The results highlight that the inference-stage architecture is just as critical as the learning rule itself for achieving the precision required for real-world robotics.