A fully convolutional neural network was applied to prestack 3D seismic data to predict lithofacies. A key challenge was to segregate volcanics and shale, which have similar elastic properties. We achieved reasonable lithofacies prediction by using pseudowells as labeled data, which could incorporate geologic information into the model (i.e., vertical distribution patterns). In particular, the addition of stratigraphic information as input to the neural network enabled effective learning. In addition, multiple networks were separately trained to evaluate prediction uncertainty. Even in such cases, the number of pseudowells required for the robust training was much smaller than the number of traces in the 3D seismic data.