also helps people avoid illness through improving their
The Bayesian perspective is a very general approach
to child development that requires specific hypoth-eses to test developmental theories (Gopnik and
Bonawitz 2014). Its principle advantage is that it allows theories to be formulated in precise and transparent ways.
The connectionists have also addressed the problem
of modeling similarity but from a different perspective
than the rational analysis used by Anderson (1990)
and the Bayesians. Rogers and McClelland (2014)
discuss this distinction within the framework of
Marr’s (1982) levels of analysis. The rational level
corresponds to Marr’s computational level, which
focuses on an analysis of the problem, including
mathematical methods to solve it. In contrast, the
connectionist approach considers how the brain —
neurons and their connections — constrain the nature of the solutions as formulated by Rumelhart,
Hinton, and McClelland (1986).
The TRACE model of auditory word recognition
(McClelland and Elman 1986) is a typical example.
The model contains three layers to represent the
temporal dynamics of word recognition by taking as
input ( 1) auditory features such as voiced and acute ( 2)
are recognized incrementally by increasing the activation
level of the correct phoneme and word units. Activation
occurs in parallel across these units and includes top-down processing that enables activation at the word level
to influence activation at the phoneme level. The top-down activation provides the same constraints on
the recognition of phonemes as the interactive activation model provides on the recognition of letters.
Another connectionist model—a simple recurrent
network—learns the semantic and syntactic properties of words by attempting to predict the next word in
a sentence (Elman 2004). The network uses each new
word in a sentence (the input) to predict the next
word in the sentence (the output). Learning occurs by
comparing the prediction with the actual occurrence.
A key step in the recurrent network is that the hidden-unit weights depend on the context unit from the
previous word, which also depended on the context
unit of the previous word, so a history of previous
information is preserved.
The similarity among the words used in the study
resulted in easily interpretable clusters based on
their hidden-unit activation patterns (Elman 2004).
The initial split partitioned the words into verbs and
nouns. The verbs are clustered into transitive verbs
and intransitive verbs. Transitive verbs, such as like
and chase, take a direct object, whereas intransitive
verbs, such as think and sleep, do not. The nouns are
clustered into animates and inanimates, which are
Figure 2. Two Categories of Schematic Faces.
From Reed and Friedman (1973). Reproduced with permission from Springer Nature, ©1973 the Psychonomic Society.