Speeding up deep neural architecture search for wearable activity recognition with early prediction of converged performance

Neural architecture search (NAS) has the potential to uncover more performant networks for human activity recognition from wearable sensor data. However, 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 over a naive approach taking the ranking of the DNNs at an early epoch as an indication of their ranking on convergence. We apply this to the optimization of the convolutional feature extractor of an LSTM recurrent network using NAS with deep Q-learning, optimizing 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 DeepConvLSTM and 0.8% better F1 score than our implementation of state-of-the-art model Attend and Discriminate, while reducing the search time by more than 90% over a random search. This opens the way to rapidly search for well-performing dataset-specific architectures. We describe the computational implementation of the system (software frameworks, computing resources) to enable replication of this work. Finally, we lay out several future research directions for NAS which the community may pursue to address ongoing challenges in human activity recognition, such as optimizing architectures to minimize power, minimize sensor usage, or minimize training data needs.