General Knowledge Emerges in Recurrent Memory Systems
Episodic memory supports the retrieval of richly detailed event memories from sparse cues. However, an event can be detailed and also lead to generalization. To explain both abilities, researchers have proposed multiple memory systems or distinct memory traces for specific and general content. In contrast, we argue that a) specificity and generalization are ubiquitous properties of any memory system, and b) powerful generalization abilities emerge when models are allowed to be dynamic and recurrent. To do this, we present the ATHENA framework, a formal framework that shows how cognitive assumptions about storage and retrieval translate into constraints on recurrent network implementations of a memory system. We then show that the resulting memory networks exhibit properties often reserved for general memory and higher-order cognition, even when only specific traces are stored. These include forming and selectively accessing abstract and hierarchical representations, inferring latent states of items or the environment, and learning relationships between these latent states (such as latent dynamics and state accessibility). We support our conclusions with formal results as well as simulation studies in artificial environments. Our findings highlight the mutual facilitation and interplay of specific and general mnemonic content and demonstrate the flexibility of single-system memory retrieval.
There is nothing here yet. Be the first to create a thread.
Cite this as: