MACHINE LEARNING ASSISTED SPANDAK 

WHAT: To prioritize visual inspection by user, we build a convolution neural
network classifier. This has helped us reject 97% of false positives. We achieved this by training the ML classifier to classify dedispersed dynamic spectra and frequency-averaged pulses. Modern supervised learning methods require large amounts of training data, so we created a synthetic data set of a wide variety of mock broadband signals embedded in the background of real telescope noise and spurious RFI. A total of around 200,000 data points were simulated, in which 100,000 of these synthetic data contained mock broadband signals embedded in real telescope backgrounds, and the remaining 100,000 contained empty backgrounds.

HOW: For every example, the model takes two inputs, a 2D dynamic spectrum, and its corresponding 1D frequency-averaged time series. These inputs are fed into two separate branches of the network —what we call the spectrogram branch and the time series branch—which extract features from the inputs and are eventually concatenated together to produce one softmax prediction. We found that architecture with two convolutional layers worked best for our application, retaining the ability to recognize signals without being unnecessarily complex. A 3 × 3 kernel with a stride length of 1 ensures that the convolution operation does not “skip over” broken broadband signals. In the time series branch, the number of filters in a convolutional layer is roughly half the number of filters in the parallel convolutional layer in the spectrogram branch to prevent overfitting, since detecting whether a peak is present in a 1D signal is not a hugely complex task. Each convolutional layer is followed by a BatchNormalization layer and a ReLU activation. We use MaxPooling layers at the end of every convolutional block in the spectrogram branch to reduce the dimensionality but found that MaxPooling layers within the time series branch led to losing the peaks in the 1D signal and thereby decided to remove them. After all convolutional blocks for both branches, we use global max pooling to transform all convolutional feature maps into a 1D tensor for each example. Subsequently, the two branches of the network are fused by concatenating the outputs of the GlobalMax- Pooling layers, after which they are fed into two fully connected layers before the prediction layer. We also use Dropout layers between the fully connected layers, which are a simple method of reducing overfitting.

TRAINING: We implement our classifier in Keras 2.0.8 and TensorFlow 1.4.1, compiling the model with a binary cross-entropy loss and the Adam optimizer Because we value recall over precision, we weight the positive class 10 times as much as the negative class, thereby penalizing the network more for missing signals than for false positives. We employ several Keras callbacks to supplement model training: ModelCheckpoint, ReduceLROnPlateau, and EarlyStopping. With a batch size of 32, we trained our models using a single Nvidia Titan XP GPU. With the validation set throughout training 134 epochs, our model converged with a validation recall of 0.9897.