Fast deep neural architecture search for wearable activity recognition by early prediction of converged performance

Neural Architecture Search (NAS) has the potential to uncover more performant networks for wearable activity recognition, but a naive evaluation of the search space is computationally expensive. We introduce neural regression methods for predicting the converged performance of a Deep Neural Network (DNN) using validation performance in early epochs and topological and computational statistics. Our approach shows a significant improvement in predicting converged testing performance. We apply this to the optimisation of the convolutional feature extractor of an LSTM recurrent network using NAS with deep Q-learning, optimising the kernel size, number of kernels, number of layers and the connections between layers, allowing for arbitrary skip connections and dimensionality reduction with pooling layers. We find architectures which achieve up to 4% better F1 score on the recognition of gestures in the Opportunity dataset than our implementation of the state of the art model DeepConvLSTM, while reducing the search time by >90% over a random search. This opens the way to rapidly search for well performing dataset-specific architectures.