Information Technology Reference
In-Depth Information
On the one hand, one may be tempted to argue that in model studies on the
emergence of spike patterns more and more biological details need to be taken into
account. For instance, only few works so far take into account additional dynamical
features such as synaptic plasticity, synaptic failure and intrinsic noise or dendritic
non-linearities (e.g. [68, 106, 112]), among others. On the other hand, model re-
duction is essential to isolate potential mechanisms that underlie any hypothesis
and thus to restrict a hypothesis as strongly as possible to make it experimentally
testable. The theoretical analyses of feed-forward chains of spiking neurons already
raise many non-trivial problems, and even the idealized recurrent neural network
models considered above typically exhibit highly complex dynamics. Idealized mod-
els enable us, nevertheless, to understand mechanisms in feed-forward and even in
recurrent networks in a systematic way.
And indeed, recent studies of reduced models for instance revealed that non-
additive dendritic integration can support the propagation of synchronous spiking
activity in recurrent networks even if they are purely randomly connected and do
not contain additional feed-forward connectivity [106]. Conceptually, this consti-
tutes another cornerstone for bridging the gap between feed-forward and recurrent
perspectives because propagation of synchronous activity similarly underlies the
emergence of spike patterns in both hypotheses. At the same time, propagation
of synchronous activity in random recurrent networks might be experimentally dis-
tinguished from that along feed-forward chains: In networks containing groups of
neurons with specic feed-forward connections between them, synchronous activity
propagates along paths that are predened by that anatomy such that propagation
takes place with high probability along the same paths if the experiment is re-
peated with the same inital group of neurons synchronized. In contrast, in random
recurrent networks without specic feed-forward structures but with non-additive
dendritic features, synchrony propagates along self-organized paths that may vary
among repetitive trials of an experiments. This mechanistic dierence implies dif-
ferent types of possible patterns of precisely timed spikes and thus a dynamics that
oers an experimental distinction between the two hypotheses.
The balanced activity described in section 13.4.3 constitutes another example
where simple models of neural circuits substantially helped to understand the bio-
logical network dynamics. The balanced state was rst investigated for binary-state
neuron models [162] and the mechanism that generates its irregular spiking behav-
ior points to a collective network eect that is due to simultaneously strong and
only weakly correlated inhibitory and excitatory inputs. Due to this basic network
mechanism, such balanced activity robustly occurs across dierent systems; it is
prevalent also for biophysically more detailed models and provides the rst consis-
tent explanation of the irregularity observed in biological neural circuits. Perhaps
reduced models will similarly help to better understand the conditions under which
spike patterns emerge in feed-forward and in recurrent networks.
