SAMPL: The Spreading Activation and Memory PLasticity Model

Beau Sievers, Ida MomennejadbioRxiv 2019 · DOI

Abstract

We present the Spreading Activation and Memory PLasticity Model (SAMPL), a computational model of how memory retrieval changes memories. SAMPL restructures memory networks as a function of spreading activation and plasticity. Memory networks are represented as graphs of items in which edge weights capture the strength of association between items. When an item is retrieved, activation spreads across nodes depending on edge weights and the strength of initial activation. A non-monotonic plasticity rule, in turn, updates edge weights following activation. SAMPL simulates human memory phenomena across a number of experiments including retrieval induced forgetting, context-based memory enhancement, and memory synchronization in conversational networks. Our results have implications for theorizing memory disorders such as PTSD and designing computationally assisted conversational therapy.

Introduction

Human memory is not a static archive but a dynamic, malleable network. The Spreading Activation and Memory PLasticity (SAMPL) model provides a unified computational framework to explain how the act of retrieval itself restructures these networks. By integrating spreading activation with a non-monotonic plasticity rule, SAMPL accounts for diverse phenomena ranging from individual forgetting to collective social memory synchronization.

The Problem: A Malleable Memory Paradox

Cognitive psychology has long observed a contradiction: retrieving a memory can either strengthen related information (context-based enhancement) or weaken it (retrieval-induced forgetting). Existing models often struggle to reconcile these effects without becoming computationally prohibitive or overly specialized. SAMPL addresses this by proposing a simple, scale-invariant principle that operates across individual and social memory structures.

Methods: The SAMPL Framework

SAMPL represents memory as a graph where nodes are items (words, concepts, or facts) and edges represent associative strength. To ensure ecological validity, the researchers initialized graphs using Word2vec semantic embeddings from the Google News Corpus.

The model operates through a four-stage cycle:

The four stages of the SAMPL model: Retrieval, Spreading Activation, Non-monotonic Plasticity, and the resulting End State.
  1. Retrieval (T1): A specific node is "fired," initiating the cycle.
  2. Spreading Activation (T2): Activation flows to neighbors, decaying as it moves further from the source (controlled by a discount parameter $\gamma$).
  3. Non-Monotonic Plasticity (T3): This "secret sauce" follows a U-shaped curve. High activation strengthens edges ("rich get richer"), while moderate activation weakens them (the "suppressive dip"). Low activation results in no change.
  4. End State (T4): The network is permanently restructured based on the plasticity updates.
Visual evolution of a memory network through the SAMPL process, illustrating the transition from initial retrieval to a restructured graph.

Key Insights and Results

The most striking insight is that monotonic plasticity (linear strengthening) cannot explain human memory. Without the "suppressive dip," the model fails to replicate forgetting.

SAMPL was validated against landmark behavioral studies:

  • Retrieval-Induced Forgetting (RIF): The model successfully simulated the "RP-" effect (ABB, 1994), where practicing some items in a category causes the suppression of unpracticed related items.
  • Contextual Dynamics: By treating context as standard nodes, SAMPL correctly predicted that repeating a context before a target enhances memory, while repeating it after does not (SHS, 2013).
  • Social Synchronization: In multi-agent simulations (CMDG, 2016), the model matched human data showing that "on-topic" conversations cause agents' memory graphs to converge, especially among those with fewer social "hops."

Discussion and Future Work

SAMPL suggests that NMP acts as a "housekeeping" rule, pruning moderately relevant associations to sharpen memory boundaries. While the current model uses bidirectional edges, future versions could incorporate hierarchical structures and asymmetric associations to better mirror hippocampal function.

Furthermore, SAMPL offers a promising path for computational psychiatry. By modeling "forgetting attacks"—using neutral cues to trigger the suppressive dip on intrusive memories—it could inform the design of computationally-assisted therapies for disorders like PTSD.

Summary updated

Abstract

Abstract We present the Spreading Activation and Memory PLasticity Model (SAMPL), a computational model of how memory retrieval changes memories. SAMPL restructures memory networks as a function of spreading activation and plasticity. Memory networks are represented as graphs of items in which edge weights capture the strength of association between items. When an item is retrieved, activation spreads across nodes depending on edge weights and the strength of initial activation. A non-monotonic plasticity rule, in turn, updates edge weights following activation. SAMPL simulates human memory phenomena across a number of experiments including retrieval induced forgetting, context-based memory enhancement, and memory synchronization in conversational networks. Our results have implications for theorizing memory disorders such as PTSD and designing computationally assisted conversational therapy.
BibTeX
@misc{sievers2019sampl,
  author = {Beau Sievers and Ida Momennejad},
  title = {{SAMPL: The Spreading Activation and Memory PLasticity Model}},
  year = {2019},
  doi = {10.1101/778563},
  howpublished = {bioRxiv},
}
Figure