Robust Model Assessment for Neural Networks

We present a robust model assessment method for neural-network models and demonstrate the approach for feedforward networks. In order to assess a neural-network model, the location of weight vectors after repeated learning experiments is analysed in the eeective weight space. We argue that appropriate or underdetermined models clearly show a cluster structure in the eeective weight space whereas too complex models do not exhibit a proper cluster structure. We achieve robust model assessment by restricting estimates of network performance to clusters.