E-DARTS: Enhanced Differentiable Architecture Search for Acoustic Scene Classification

Neural architecture search (NAS) has attracted a lot of attention in recent years due to the growing interest in automating the neural architecture design. Recently, Differentiable Architecture Search (DARTS) has achieved promising results on the image classification task with a significantly reduced search cost to only a few GPU days. Specifically, DARTS relaxed the search space to be continuous and thus allowing the architecture to be optimized with gradient descent, where a set of simple candidate operations were used to construct the searched network. However, in our experiments, original DARTS achieved a limited classification accuracy of 50.45 ± 2.34% on the acoustic scene classification (ASC) dataset of DCASE2020 Task lA. In this work, we propose an enhanced version of DARTS, namely Enhanced-DARTS (E-DARTS), which uses residual and squeeze-excitation blocks to build more complex networks with DARTS. The aforementioned blocks were used as candidate operations for the super-network in addition to the convolutional and pooling layers used in original DARTS. The proposed approach has achieved a significantly improved classification accuracy of 64.57 ± 0.87% on the same dataset.