A study of sample size with neural network

This study investigated sample complexity for a linearly separable dataset by training and testing a breast cancer database. This study considered two networks: a single layer network and a multilayer network. We observed that the training sample size could be 1 for both networks with good generalization results under different conditions. The multilayer network performed well with any training sample but the single layer network required selection of a training sample having informative class output value. When the multilayer network was trained with a small training sample and the threshold for the testing network output was set at an appropriate value the test error became as low as 2%. We concluded that for a linearly separable dataset it is possible achieve good performance by training a network with small sample size, such as 1 or 2.