AMG: a memory generation model based on Hebbian plasticity and inferential association

The ability to extract ambiguous memories from past experiences is crucial for cognitive activities in the human brain. This capability is equally important for artificial intelligence agents (AI Agents). Implementing generative models to facilitate the extraction of ambiguous memories can not only enhance the cognitive abilities of AI Agents but also has a positive significance in real life, such as assisting Alzheimer's patients in recovering lost memories from limited cues. This paper proposes a memory association model, named generative associative memory (AMG), which combines Hebbian plasticity and Conditional Deep Convolutional Generative Adversarial Network (CDCGAN). The striking feature of AMG is the ability of simulating the process of memory completion in the human brain. The distinctive idea of AMG is twofold. One is associative memory. The feature vectors of the memories are extracted and stored as key-value pairs within an association matrix constructed based on Hebbian plasticity. The other is inference generation. The association vectors are inferred from the association matrix and used as conditional inputs for CDCGAN, enabling the model to effectively reconstruct the missing components, thereby simulating the memory recovery process observed in the human brain. Experimental results demonstrate that AMG effectively recovers residual memory, achieving a structural similarity index of 0.9951, outperforming non-Hebbian models and reducing total runtime by 9.06\%.