Low-frequency earthquakes are a seismic manifestation of slow fault slip. Their emergent onsets, low amplitudes, and unique frequency characteristics make these events difficult to detect in continuous seismic data. Here, we train a convolutional neural network to detect low-frequency earthquakes near Parkfield, CA using the catalog of Shelly (2017), https://doi.org/10.1002/2017jb014047 as training data. We explore how varying model size and targets influence the performance of the resulting network. Our preferred network has a peak accuracy of 85% and can reliably pick low-frequency earthquake (LFE) S-wave arrival times on single station records. We demonstrate the abilities of the network using data from permanent and temporary stations near Parkfield, and show that it detects new LFEs that are not part of the Shelly (2017), https://doi.org/10.1002/2017jb014047 catalog. Overall, machine-learning approaches show great promise for identifying additional low-frequency earthquake sources. The technique is fast, generalizable, and does not require sources to repeat.