Critical Neuron

"Critical neurons" in neural networks are those neurons deemed most influential in a model's function, whether for accurate classification, memorization of training data, or vulnerability to adversarial attacks. Current research focuses on identifying and manipulating these neurons across various architectures, including large language models, diffusion models, and spiking neural networks, using techniques like neuron pruning, noise injection, and attribution methods to improve model efficiency, robustness, and interpretability. This research is significant for enhancing the reliability and trustworthiness of AI systems in safety-critical applications and for gaining a deeper understanding of how these complex systems learn and function.

Papers