Subsymbolic behavior
History
Some properties
- Representations are often grounded: there is a straightforward path between perceptual input, internal representations, and motor output.
-
Behaviors may be emergent rather than represented explicitly.
- Corresponding to the atomic symbol tokens of symbolic models, there may be (non-atomic) patterns (vectors).
The primitive elements of these patterns may be thought of as
"neurons" or "processing units" in neural networks.
- Corresponding to the structured representations of symbolic models are just more patterns of the same type (and length) as the "symbol" patterns.
- Representations are thus distributed, in the most extreme cases, holographic.
- Input-output behavior, when it is part of the model,
is governed entirely by associations.
In the most radical, dynamical systems models, there is no real input
and output.
-
Similarity between representations is often vector distance.
- Processes usually happen in parallel.
- There is usually no central controller which
coordinates the activities of the modules of the cognitive
system and
selects among candidate processes at each point in time.
- There may be no distinction between the cognitive system and
systems for perception and action.
- Knowledge is usually learned, through some sort of associationist process, rather than programmed into the system.
- Time may be taken seriously.
Examples
- A theory of "meaning" in which a word's meaning is based entirely on its pattern of co-occurrences with other words:
Latent Semantic Analysis
- Create a massive word-document co-occurrence matrix (list of lists in Python) for all of the words in a large corpus in some language.
Reduce the dimensionality of the matrix.
Represent the "meaning" of each word as its resulting vector.
-
A passive English output sentence is generated one word at a time, given a constant input pattern representing an agent, an action, a patient, and focus on the patient.
This is accomplished using a type of recurrent neural network in which the output layer of processing units represents a single word and a separate layer of units stores the current state, that is, a representation of what has been produced so far.