H. Wersing et al., Dynamical stability conditions for recurrent neural networks with unsaturating piecewise linear transfer functions, NEURAL COMP, 13(8), 2001, pp. 1811-1825
Citations number
24
Categorie Soggetti
Neurosciences & Behavoir","AI Robotics and Automatic Control
We establish two conditions that ensure the nondivergence of additive recur
rent networks with unsaturating piecewise linear transfer functions, also c
alled linear threshold or semilinear transfer functions. As Hahnloser, Sarp
eshkar, Mahowald, Douglas, and Seung (2000) showed, networks of this type c
an be efficiently built in silicon and exhibit the coexistence of digital s
election and analog amplification in a single circuit. To obtain this behav
ior, the network must be multistable and nondivergent, and our conditions a
llow determining the regimes where this can be achieved with maximal recurr
ent amplification. The first condition can be applied to nonsymmetric netwo
rks and has a simple interpretation of requiring that the strength of local
inhibition match the sum over excitatory weights converging onto a neuron.
The second condition is restricted to symmetric networks, but can also tak
e into account the stabilizing effect of nonlocal inhibitory interactions.
We demonstrate the application of the conditions on a simple example and th
e orientation-selectivity mo del of Ben-Yishai, Lev Bar-Or, and Sompolinsky
(1995). We show that the conditions can be used to identify in their model
regions of maximal orientation-selective amplification and symmetry breaki
ng.