On the Infeasibility of Training Neural Networks with Small Squared Errors
–Neural Information Processing Systems
We demonstrate that the problem of training neural networks with small (average) squared error is computationally intractable.
Neural Information Processing Systems
Dec-31-1998