-
Notifications
You must be signed in to change notification settings - Fork 0
0.1 Distributed memory
Patterns that are scaled by a network are called eigenvectors, eigenvector simply means "same vector." The magnitude of the eigenvector, as it is processed by the network, is called its eigenvalue.
The view that human memory is physiologically distributed within circumscribed regions of the brain seems to be quite a reasonable and plausible assumption.
But given the rather loose coupling between a psychological or cognitive theory of physiological implementation, we can ask, does the notion of distributed memory have anything to offer us in terms of an understanding of human cognition?
Our model shares a number of basic assumptions about the nature of the processing and memory system with most other distributed models.
In particular, the processing system is assumed to consist of a highly interconnected network of units that take on activation values and communicate with other units by sending signals modulated by weights associated with the connections between the units. Sometimes we may think of the units as corresponding to particular representational primitives, but they need not. For example, even what we might consider to be a primitive feature of something, like having a particular color, might be a pattern of activation over a collection of units.
We assume that the units are organized into modules. Each module receives inputs from other modules; the units within the module are richly interconnected with each other; and they send outputs to other modules. The state of each module represents a synthesis of the states of all the modules it receives inputs from.
Others will come from relatively more abstract modules, which themselves receive inputs from and send outputs to other modules placed at the abstract end of several different modalities. Thus, each module combines a number of different sources of information.
A pattern of activation only counts as the same as another if the same units are involved.
-
State as pattern of activation; In a distributed memory system, a mental state is a pattern of activation over the units in some subset of the modules. The patterns in the different modules capture different aspects of the content of the states in a kind of a partially overlapping fashion. Alternative mental states are simply alternative patterns of activations over the modules. Information processing is the process of evolution in time of mental states (?).
-
Memory traces as changes in the weights; Patterns of activation come and go, leaving traces behind when they have passed. What are the traces? They are changes in the strengths or weights of the connections between the units in the modules. Each memory trace is distributed over many different connections, and each connection participates in many different memory traces. The traces of different mental states are therefore superimposed in the same set of weights.
-
Retrieval as reinstatement of prior pattern of activation; Retrieval amounts to partial reinstatement of a mental state, using a cue which is a fragment of the original state. For any given module, we can see the cues as originating from outside of it. Some cues could arise ultimately from sensory input. Others would arise from the results of previous retrieval operations, feedback to the memory system under the control of a search or retrieval plan.