Biologically-Detailed Network Modelling - Computational Neuroscience: A Comprehensive Approach - page 298

The first necessary step is to test that the model can reproduce the experimental
results that were used in developing the model, the 'built-in' properties. This test
is important first to check for mistakes in the implementation of the model (such as
'bugs' in the code) but can also expose inconsistencies between different sources of
experimental data. This alone is not sufficient, however. It is essential to attempt to
reproduce experimental results that were not used in developing the model. If the
'emergent' properties of the model match experimental findings, this validates the
model. (For more discussion on built-in vs emergent properties of computational
models, see Protopapas et al. (1998) [29]).

As an example, one way in which we tested our olfactory bulb model was by
attempting to reproduce published data on dendrodendritic synaptic currents [34].
Many of the model parameters were derived from this same publication, but the am-
plitude and time constant of the mitral cell IPSC were not incorporated directly in the
model: they are emergent properties. The simulated IPSC matched the experimental
one closely, although with a number of discrepancies. Because of these discrepan-
cies we would regard the model as only partially validated by these results. The

resolution of these discrepancies suggests further lines of enquiry.

10.8 Conclusions

What is the future of biologically-detailed network modelling? As computers be-
come more powerful, more detail can be incorporated into models, but with this
comes the need for more detailed, carefully-designed biological experiments to con-
strain and test the models. As the complexity of models approaches that of real
systems, more sophisticated analysis tools will be required. As discussed in the In-
troduction, it will in most cases be desirable to develop a hierarchy of models to
link abstract, conceptual models, via models of intermediate complexity, to detailed
models and thence to biological data.

References
[1] G. A. Ascoli (1999), Progress and perspectives in computational neuroanatomy,

Anatomical Record, 257 : 195-207.

[2] G. A. Ascoli, and J. L. Krichmar (2000), L-Neuron: a modeling tool for the ef-

ficient generation and parsimonious description of dendritic morphology, Neu-
rocomputing , 32-33 : 1003-1011.