phology in a coherent model reduction framework has come in the past five years.
Drawing on tools from computational physics, Woo, Yang and Choi demonstrated
that a passive cable can be simulated using an eigenfunction expansion approach us-
ing more than an order of magnitude fewer ordinary differential equations (ODEs)
and having greater accuracy than the compartmental approach (Woo et al., 2005).
Woo and Choi subsequently extended this work to deal with myelinated axons (Woo
and Choi, 2007), and, using a pseudo-spectral method, Shin, Yang, and Choi further
extended these ideas to work with active cables (Shin et al., 2009).
While in each case these methods produced more accurate results with fewer
ODEs, the work raises some questions. First, the authors present no timing com-
parisons between the full and reduced models, making it hard to determine how
computationally efficient this reduction really is. Second, the active cable model used
in (Shin et al., 2009) is not a conductance-based model, but a simpler system used
by Wilson (Wilson, 1999). Finally, and most importantly, the authors consider input
patterns only to single locations along the cable. Although they vary the stimulus
location for different examples, they do not give results for simultaneous spatially-
distributed inputs, and in fact the dimension of the reduced models appears to depend
significantly on the stimulus location. Despite these criticisms, the work from Choi’s
group demonstrates that nonlinear model reduction is possible in simple cases and
bolsters our conviction that drastically reduced models can be found for realistic
neuronal models.