83
branch and yields the final snapshot set (see Figure 3.3). Note that since we have two
snapshot sets, one for voltage and one for the nonlinear terms, the tolerances εgιobai
and ειocaι can be different for each set.
3.3.3 Forked Neuron Results
Equipped with the tools from the above sections, we return to the task of reducing
the forked neuron. We take 200 snapshots over a 10 ms window (using ∆t = 0.1 ms),
and we obtain them by giving a suprathreshold stimulus to a distal branch. We
use the branchwise orthogonalization algorithm to generate snapshots of branches
in isolation, where the active zone is computed by applying V-Slim with tolerances
ɛ focal = 0> εk>cal = 0> ɛglobaɪ = 1θ^^6> and ɛglobaɪ = 10^5> and then saving everY 4th
snapshot from this resulting set. Simulations consist of 750 random step currents
injected over a 1 second interval, with each current having amplitude 0-60 pA and
lasting 0-5 ms.
Branch-Ortho turns out to be a very effective method of improving the accuracy
of the reduced system, as Table 3.3 demonstrates. Not only are spike times accu-
rately reproduced, but the somatic voltage traces are nearly exactly duplicated in the
reduced system, as shown in Figure 3.5. The improvement can be seen qualitatively
in Figure 3.4 by observing that the DEIM points are more evenly spaced throughout
the neuron than they are without Branch-Ortho.
Our next test is done with the HHA model, but with all the other parameters