The experience gained in many experiments with neural networks has shown that many challenging problems are still hard to solve, since the learning process becomes very slow, often leading to sub-optimal solutions. In this paper we analyze this problem for the case of two-layered networks by discussing on the joint behavior of the algorithm convergence and the generalization to new data. We suggest two scores for generalization and optimal convergence that behave like conjugate variable in Quantum Mechanics. As a result, the requirement of increasing the generalization is likely to affect the optimal convergence. This suggests that 'difficult' problems are better face with biased-models, somewhat tuned on the task to be solved.