On minimizers and convolutional filters: a partial justification for the unreasonable effectiveness of CNNs in categorical sequence analysis.