We study the use of feed-forward convolutional neural networks for the unsupervised problem of mining recurrent temporal patterns mixed in multivariate time series. Traditional convolutional autoen-coders lack interpretability for two main reasons: the number of patterns corresponds to the manually-fixed number of convolution filters, and the patterns are often redundant and correlated. To recover clean patterns, we introduce different elements in the architecture, including an adap-tive rectified linear unit function that improves patterns interpretability, and a group-lasso regularizer that helps automatically finding the relevant number of patterns. We illustrate the necessity of these elements on synthetic data and real data in the context of activity mining in videos.