Wide attention was recently given to the problem of fault-tolerance in neural networks; while most authors dealt with aspects related to specific VLSI implementations, attention was also given to the intrinsic capacity of survival to faults characterizing the neural modes. The present paper tackles this second theme, considering in particular multilayered feed forward nets. One of the main goals is to identify the real influence of faults on the neural computation in order to show that neural paradigms cannot be considered intrinsically fault tolerant (i.e., able to survive to faults, even several of the most common and simple ones). A high abstraction level (corresponding to the neural graphs) is taken as the basis of the study and a corresponding error model is introduced. The effects of such errors induced by faults are analytically derived to verify the probability of intrinsic masking in the final neural outputs. Then, conditions allowing for complete compensation of the errors induced by faults through weight adjustment are evaluated to test the masking abilities of the network. The designer of a neural architecture should perform such a mathematical analysis to check the actual fault-tolerance features of his or her system. Unfortunately, this involves a very high computational overhead. As a cost-effective alternative for the designer, the use of a behavioral simulation is proposed for a quantitative evaluation of the error effect on the neural computation. Repeated learning (i.e., a new application of the learning procedure on the faulty network) is then experimented to induce error masking. Experimental results prove that even single errors affect the computation in a relevant way and that weight redistribution is not able to induce complete masking after a fault occurred, i.e., the network cannot be considered per se intrinsically fault tolerant and it is not possible to rely on learning only in order to achieve complete masking abilities. Mapping criteria of physical faults onto the abstract errors are finally examined to show the usability of the proposed analysis in evaluating the actual robustness of a neural networks' implementation and in identifying the critical areas where architectural redundancy should be introduced to achieve fault tolerance.