Page 206 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

196 CHAPTER 9 Theory of the Brain and Mind: Visions and History

4. IS COMPUTATIONAL NEUROSCIENCE SEPARATE FROM
NEURAL NETWORK THEORY?

The growth of the neural modeling field in the 1990s and early 2000s led to the for-
mation of different research niches. Since then, an increasing number of researchers
have taken up the challenge of modeling complex data from neuropsychology and
cognitive neuroscience. Many of these researchers seldom if ever attend IJCNN
conferences or publish in Neural Networks, and they are not part of the INNS
community (although some of them attend other conferences that cut across
engineering and neuroscience concerns, such as Neural Information Processing
Systems or NIPS).
In particular, computational neuroscience since the 1990s has become a field
with its own meetings that are largely disconnected from neural network or artificial
intelligence meetings. The concern of many computational neuroscientists has been
about processes at the level of neurons and small neural ensembles without relating
those processes to cognitive or behavioral functions (see, e.g., Ref. [55]). Yet more
recently Ashby and Helie [56] have proposed that the best computational models of
brain function are based on a synthesis of these fields that they call computational
cognitive neuroscience:
Computational Cognitive Neuroscience (CCN) is a new field that lies at the inter-
section of computational neuroscience, machine learning, and neural network
theory (i.e., connectionism).

The advantages that Ashby and He ´lie posit for CCN models as opposed to purely
cognitive models include the ability to make predictions about neural processes such
as fMRI, EEG, and single-unit responses, as well as behavioral data. Yet CCN is really
not new but has been a gradual development over 20e30 years, in part as an
outgrowth of previous neural models that were more abstract and cognitive and had
less neuroscientific detail. The third edition of my textbook [1] uses “Computational
Cognitive Neuroscience” as the title for its second half and emphasizes the roots of
those models in the more introductory models of the book’s first half.
Current models which can be considered CCN come from a variety of intellectual
sources. One of those sources is the back propagation algorithm. One of the earliest
models in this tradition was a network theory of the Stroop task, which involves
naming the color of ink in a word that represents either the same color or a different
color [57,58]. This model included a simulation of poor performance on this task by
schizophrenics, based on “lesioning” of a particular neural network connection.
Subsequently O’Reilly [59] set out to mimic the effects of the back propagation
algorithm using the LEABRA equations which were purportedly more biologically
realistic.
O’Reilly [60] followed up this work with proposing six principles for computa-
tional modeling of the cortex. This approach led to a sophisticated set of interrelated
models which simulated such processes as working memory [61,62] and Pavlovian