On the complexity of loading shallow neural networks (Q1105389)

From MaRDI portal
scientific article
Language Label Description Also known as
English
On the complexity of loading shallow neural networks
scientific article

    Statements

    On the complexity of loading shallow neural networks (English)
    0 references
    1988
    0 references
    We formalize a notion of loading information into connectionist networks that characterizes the training of feed-forward neural networks. This problem is NP-complete, so we look for tractable subcases of the problem by placing constraints on the network architecture. The focus of these constraints is on various families of ``shallow'' architectures which are defined to have bounded depth and unbounded width. We introduce a perspective on shallow networks, called the support cone interaction (SCI) graph, which is helpful in distinguishing tractable from intractable subcases: When the SCI graph is a tree or if of limited bandwidth, loading can be accomplished in polynomial time; when its bandwidth is not limited we find the problem NP-complete even if the SCI graph is a simple 2-dimensional planar grid.
    0 references
    0 references
    0 references
    0 references
    0 references
    support cone interaction graph
    0 references
    loading
    0 references
    connectionist networks
    0 references
    neural networks
    0 references
    NP-complete
    0 references
    shallow networks
    0 references
    0 references