Clustering is an important tool in data mining and knowledge discovery. Fuzzy Adaptive Resonance Theory, a member of unsupervised neural networks, clusters data effectively because of applying operator AND of Fuzzy Logic. In previous studies, learning from data was ineffective when the surface of data is higher than the surface of weight vector of categories. In this paper, we propose an improved learning rule to learn from data better. In the proposed rule, the weights of wining category are decreased to adapt to each input. Each input shows the effect for categories by a learning parameter. The learning parameter is adjusted until the best stable state and performance of clustering results are achieved. We have conducted experiments on 10 benchmark datasets to prove the effectiveness of the proposed rule. The experiment results showed that Fuzzy ART learned with the improved rule (IFART-Improved Fuzzy Adaptive Resonance Theory) performs better than existing models in complex small datasets.