This research aims to predict knock occurrences by deep learning using in-cylinder pressure history from experiments and to elucidate the period in pressure history that is most important for knock prediction. Supervised deep learning was conducted using in-cylinder pressure history as an input and the presence or absence of knock in each cycle as a label. The learning process was conducted with and without cost-sensitive approaches to examine the influence of an imbalance in the numbers of knock and non-knock cycles. Without the cost-sensitive approach, the prediction accuracy exceeded 90% and both the precision and the recall were about 70%. In addition, the trade-off between precision and recall could be controlled by adjusting the weights of knock and non-knock cycles in the cost-sensitive approach. Meanwhile, it was found that including the pressure history of the previous cycle did not influence the classification accuracy, suggesting little relationship between the combustion behavior of the previous cycle and knock occurrence in the following cycle. Moreover, learning the pressure history up to 10° CA before a knock improved the classification accuracy, whereas learning it within 10° CA before a knock did not noticeably affect the accuracy. Finally, deep learning was conducted using data, including multiple operating conditions. The present study revealed that deep learning can effectively predict knock occurrences using in-cylinder pressure history.