A new feature evaluation index based on fuzzy set theory and a connectionist model for its evaluation are provided. A concept of flexible membership function incorporating weighting factors, is introduced which makes the modeling of the class structures more appropriate. A neuro-fuzzy algorithm is developed for determining the optimum weighting coefficients representing the feature importance. The overall importance of the features is evaluated both individually and in a group considering their dependence as well as independence. Effectiveness of the algorithms along with comparison is demonstrated on speech and Iris data.