Complex-valued temporal radio signals are used as input with minimal pre-processing to a neural network. Expert feature extraction is avoided, but instead, truncated frames consisting of in-phase and quadrature (I/Q) data samples are supplemented with their Discrete Fourier Transforms of equal length and the continuous finite phase difference to produce inputs to the neural network. Based on the experiment, it is demonstrated that a convolutional neural network can accurately classify many common modulation types without the use of phase-locked loops or the need to calculate additional statistical quantities or higher-order moments. The classification accuracy of some modulation schemes, notability frequency shift keyed types, was shown to be over 95% at SNR greater than 10 dB. Conversely, PSK and QAM-type modulations that incorporate the instantaneous phase as a symbol component are demonstrated more challenging for this model to classify above 85% accuracy.