Dynamics of Memory: Encoding and Representation in the Brain
Published:
How do we encode information? How are memories represented in the brain? Do we value each episode of information, or do we integrate new information into an existing abstract world model? Behavioral data support various theories. This paper introduces two computational models of memory formation and explores their similarities in the mechanisms for updating dynamic cognition.
Distributed Memory
Proposed by McClelland & Rumelhart in 1985, the distributed memory model posits that information is stored across a network. The prototype is essentially a linear combination of specific exemplars. While unique events may bias perception, generally, prototypes guide our interpretation of the world. Memory abstracts the central tendency from a series of disparate experiences, affording minimal weight to the idiosyncrasies of those experiences. Key Properties of the Distributed Memory Model:
- Simple, Highly Interconnected Units: Units do not correspond one-to-one with physical entities. For instance, colors may be represented by patterns of activation across various units.
- Modular Structure: Units are organized into modules, each receiving input from and sending output to other modules. These modules are highly interconnected and can represent different types of information, allowing for diverse combinations rather than mere linear aggregation.
- Mental States as Activation Patterns: Mental states are reflected by different activation patterns within these modules. New information modifies the strength or weight of connections, accounting for observed behavioral patterns like familiarity effects. This model is efficient, negating the need for vast storage capacity and complex search mechanisms. It also suggests a seamless continuity between episodic and semantic memory, with the latter arising from the superposition of episodic traces.
Multiple-Trace Memory Model
In contrast, Hintzman’s 1988 model introduces a more intricate system, distinguishing episodic memory from generic or semantic memory. It envisions long-term memory as a collection of episodic traces. Encoding involves copying an event into long-term storage, with retrieval triggered by cues that activate all traces simultaneously, creating a composite echo. Semantic abstraction is facilitated through inter-trace resonance, where activated traces can amplify each other, akin to the interconnected nodes in the distributed model.
Another multi-model theory of memory, proposed by Raaijmakers and Shiffrin, suggests that long-term memory consists of a richly interconnected network with various levels, strata, categories, and hierarchical structures encompassing a diverse array of relationships, schemas, frames, and associations. This stands in contrast to the distributed model of memory, which posits that memories are formed from the combinations of different nodes within a network.
In the associative memory model, the elements of memory are often referred to as “images.” When these images are associated with temporal-contextual information, they become equivalent to an episode as defined by Tulving in 1972 or as a token akin to Anderson and Bower’s 1973 concept. However, some images may lack a specific contextual basis; these are typically known as semantic images or types.
The theory posits that when the contextual coding is weak or absent, individuals are likely to recall only the images devoid of context. For instance, in tasks where meaning is emphasized over the context, the additional information stored alongside a word in a new image might be semantic rather than temporally or situationally specific.
To qualify as an image in this model, it must have the characteristic that when it is accessed during a memory search, the recovery of the available portion of its informational content occurs directly, without the need for additional memory searches using the same probe cues. Essentially, this information becomes immediately accessible within working memory.
##Reconciling Models: Optimal Filtering
Gershman et al. (2014) suggest reconciling these models by considering the environment’s influence on memory systems. Memory structures might reflect optimal filtering in a dynamically changing environment, encoding distinct “dynamical modes.” New models are inferred when abrupt discontinuities in sensory data occur, which existing models cannot explain. This leads to either modification of an existing memory or the formation of a new one.
References
- Gershman, S. J., Radulescu, A., Norman, K. A., & Niv, Y. (2014). Statistical Computations Underlying the Dynamics of Memory Updating. PLoS Computational Biology, 10(11), e1003939. https://doi.org/10.1371/journal.pcbi.1003939
- Hintzman, D. L. (1988). Judgments of frequency and recognition memory in a multiple-trace memory model. Psychological Review, 95(4), 528–551. https://doi.org/10.1037/0033-295X.95.4.528
- McClelland, J. L., & Rumelhart, D. E. (1985). Distributed memory and the representation of general and specific information. Journal of Experimental Psychology: General, 114(2), 159–188. https://doi.org/10.1037/0096-3445.114.2.159
- Raaijmakers, J. G. W., & Shiffrin, R. M. (1981.). Search of Associative Memory.