AI Engram: In Search of Memory Traces in Artificial Intelligence

The paper introduces a geometric framework for identifying "AI engrams," which are memory traces in deep neural networks, by formalizing neuroscientific criteria into a constrained inverse problem. It presents a closed-form estimator that allows for the isolation of individual memory traces from entangled parameters and demonstrates the ability to manipulate learned knowledge through linear arithmetic, applicable across various models from MLPs to LLMs. This research provides insights into the relationship between biological memory and artificial representation learning, highlighting the potential for more efficient memory management in AI systems.