In the classification of electroencephalograms for a brain-computer interface (BCI), two steps are generally applied: preprocessing for feature extraction and classification using a classifier. As a result, combinations of a myriad of preprocessing and a myriad of classification method have disordered for each classification target and data. Conversely, neural networks can be applied to any classification problem because they can transform an arbitrary form of input into an arbitrary form of output. We considered a transposed convolution as a preprocessor that can set the window width and number of output features and classified it using a convolutional neural network (CNN). Using a simple CNN with a transposed convolution in the first layer, we classified the data of the motor imagery tasks of the BCI competition IV 2 dataset. The results showed that, despite not being among the best conventional methods available, we were still able to obtain a high degree of accuracy.