Do Spikes Protect Privacy? Investigating Black-Box Model Inversion Attacks in Spiking Neural Networks

A new study introduces Temporal-adaptive Weight Quantization (TaWQ) for Spiking Neural Networks (SNNs), which aims to reduce energy consumption while maintaining accuracy. This method leverages temporal dynamics to allocate ultra-low-bit weights, demonstrating minimal quantization loss of 0.22% on ImageNet and high energy efficiency in extensive experiments.

A new study has proposed an efficient Path Integration (PI) approach that utilizes representation learning models to replicate the neurodynamic patterns of Continuous Attractor Neural Networks (CANNs). This method successfully reconstructs Head Direction Cells (HDCs) and Grid Cells (GCs) using lightweight Artificial Neural Networks (ANNs), enhancing the operational efficiency of Brain-Inspired Navigation (BIN) technology.

Recent research has demonstrated that random spiking neural networks (SNNs), particularly leaky integrate-and-fire (LIF) models, exhibit stability and a concentration of their Fourier spectrum on low-frequency components, which enhances their performance in classification tasks. This study introduces the concept of spectral simplicity, linking it to the simplicity bias observed in deep networks.

MonoKAN, a Certified Monotonic Kolmogorov-Arnold Network, has been introduced to enhance the interpretability of Artificial Neural Networks (ANNs) while ensuring compliance with partial monotonicity constraints. This development addresses the ongoing challenges in achieving transparency and accountability in AI applications, particularly where model predictions must meet expert-defined requirements.