Relations between different hierarchical levels in communities of benthic macroinvertebrates were patterned by using the counterpropagation neural network. Two data sets in densities of different level in taxonomy (e. g., Family and Genus/Species) and functional groups were provided concurrently as input and output to the neural network. A "cross-training" by the counterpropagation network was conducted between two sets of data. In the trained patterns, abundant groups appeared more consistently while the groups in low densities tended to disappear. The patterned relationships between the hierarchical levels reflected the variation of community groupings and the "two-way" patterning between input and output data was possible. Through the recognition process the trained network was further able to forecast the densities at the other hierarchical layer in a time-delayed manner if previous community data were given as input.were given as input.