Publication View

For Neural Networks, Function Determines Form (1992)

Abstract
This paper shows that the weights of continuous-time feedback neural networks are uniquely identifiable from input/output measurements. Under very weak genericity assumptions, the following is true: Assume given two nets, whose neurons all have the same nonlinear activation function oe; if the two nets have equal behaviors as "black boxes" then necessarily they must have the same number of neurons and ---except at most for sign reversals at each node--- the same weights. Moreover, even if the activations are not a priori known to coincide, they are shown to be also essentially determined from the external measurements. Key words: Neural networks, identification from input/output data, control systems 1 Introduction Many recent papers have explored the computational and dynamical properties of systems of interconnected "neurons." For instance, Hopfield ([7]), Cowan ([4]), and Grossberg and his school (see e.g. [3]), have all studied devices that can be modelled by sets of nonlinear dif...

Publication details
Download http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.48.3536
Source ftp://math.rutgers.edu/pub/sontag/net-ident.ps.gz
Contributors CiteSeerX
Repository CiteSeerX - Scientific Literature Digital Library and Search Engine (United States)
Type text
Language English
Relation 10.1.1.38.4197, 10.1.1.47.8383, 10.1.1.29.6931, 10.1.1.50.8970, 10.1.1.41.1536, 10.1.1.55.8164, 10.1.1.52.1694, 10.1.1.53.4648, 10.1.1.49.6388, 10.1.1.27.8586, 10.1.1.48.5116, 10.1.1.55.8244, 10.1.1.59.2508, 10.1.1.42.3113, 10.1.1.54.4583, 10.1.1.48.7246, 10.1.1.45.7232, 10.1.1.75.6204, 10.1.1.78.9875, 10.1.1.79.440, 10.1.1.81.2954, 10.1.1.81.5605, 10.1.1.81.9902, 10.1.1.127.452, 10.1.1.50.8809, 10.1.1.40.6896, 10.1.1.26.5343, 10.1.1.26.5595, 10.1.1.17.6494, 10.1.1.7.6537, 10.1.1.5.4856