Modelling Autobiographical Memory Loss across Life Span

Di Wang, Ah-Hwee Tan, Chunyan Miao, Ahmed A. Moustafa

Introduction

In the field of Artificial Intelligence, data storage is often treated as a perfect, lossless process. However, biological systems operate differently: forgetting is not a failure of the system, but a functional feature of a self-regulating mechanism constrained by neural "hardware." While cognitive architectures like ACT-R and Soar include memory modules, they often lack a quantitative, neurocognitively grounded framework for how autobiographical memory degrades over a human lifetime. This paper introduces the Autobiographical Memory-Adaptive Resonance Theory (AM-ART), a first-of-its-kind model designed to emulate the nuanced patterns of memory loss observed in humans over a 70-year span.

The AM-ART Framework

The researchers developed a three-layer neural network architecture that maps onto specific brain regions involved in memory processing. The first layer ($F_1$) represents the temporal lobe, processing sensory and contextual inputs such as time, location, people, and emotion. The second layer ($F_2$) encodes specific events, while the third layer ($F_3$) represents the hippocampus, where these events are integrated into broader episodic patterns.

The AM-ART neurocomputational architecture, illustrating the flow from sensory inputs in the temporal lobe to episodic integration in the hippocampus.

To simulate the biological reality of forgetting, the model incorporates three distinct neural processes:

  1. Overload ($\lambda$): Occurs during the formation phase. It models cognitive overload where specific details fail to be encoded because input channels are overwhelmed.
  2. Decay ($\phi$): Occurs during storage. Following an exponential curve, the vividness of a memory fades over time unless it is reactivated or rehearsed.
  3. Inhibition ($\mu$): Occurs during retrieval. This models "retrieval-induced forgetting," where similar memories compete for access, or emotional states prevent the retrieval of an existing memory.

Methodology and Data

To provide the model with a simulated lifetime of experiences, the authors utilized the Obama Life Events Dataset, consisting of 1,019 snapshots of Barack Obama’s life. These events were distributed across eight life stages (ages 0 to 70). The model's parameters were tuned using a Genetic Algorithm (GA) to match real-world human behavioral data from landmark psychological studies by Berntsen and Rubin (2002) and Rubin and Schulkind (1997).

Results: Replicating Human Memory Patterns

The AM-ART model demonstrated a high correlation (0.793 ± 0.166) with human memory recall performance. Crucially, the model successfully replicated two key psychological phenomena:

  • The Reminiscence Bump: A disproportionately high recall of events from adolescence and early adulthood (ages 10–30).
  • The Recency Effect: Higher recall rates for events that occurred more recently in the subject's life.
Comparison of human behavioral data (a) showing memory recall across the lifespan versus the AM-ART model's emulated results (b), demonstrating the model's ability to capture the 'reminiscence bump'.

Furthermore, the model emulated human emotional bias, where pleasant memories are recalled more frequently than unpleasant ones. This was achieved through an "emotional coefficient" that regulates retrieval inhibition based on the memory's affective state. The study also found that memory performance declines more rapidly in later life stages, requiring a polynomial rather than a linear model to accurately capture aging dynamics.

Conclusion and Future Work

The AM-ART model provides a robust mathematical foundation for understanding how humans lose autobiographical information. While the study relied on a proxy dataset for a 70-year life span, its predictive power—such as using data from 40-year-olds to predict their recall performance in their 60s—is a significant step forward. Future research aims to implement these memory discard strategies in autonomous agents, potentially creating robots that prioritize recent or emotionally significant memories, making them more relatable and efficient in human-robot interactions.

Summary updated

Abstract

Neurocomputational modelling of long-term memory is a core topic in computational cognitive neuroscience, which is essential towards self-regulating brain-like AI systems. In this paper, we study how people generally lose their memories and emulate various memory loss phenomena using a neurocomputational autobiographical memory model. Specifically, based on prior neurocognitive and neuropsychology studies, we identify three neural processes, namely overload, decay and inhibition, which lead to memory loss in memory formation, storage and retrieval, respectively. For model validation, we collect a memory dataset comprising more than one thousand life events and emulate the three key memory loss processes with model parameters learnt from memory recall behavioural patterns found in human subjects of different age groups. The emulation results show high correlation with human memory recall performance across their life span, even with another population not being used for learning. To the best of our knowledge, this paper is the first research work on quantitative evaluations of autobiographical memory loss using a neurocomputational model.
BibTeX
@inproceedings{wang2019modelling,
  author = {Di Wang and Ah-Hwee Tan and Chunyan Miao and Ahmed A. Moustafa},
  title = {{Modelling Autobiographical Memory Loss across Life Span}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2019},
  doi = {10.1609/aaai.v33i01.33011368},
}
Figure