Layered Feature Representation for Differentiable Architecture Search
Abstract Differentiable architecture search (DARTS) approach has made great progress in reducing the com- putational costs of neural architecture search. DARTS tries to discover an optimal architecture module called cell from a predefined super network. However, the obtained cell is then repeatedly and simply stacked to build a target network, failing to extract layered fea- tures hidden in different network depths. Therefore, this target network cannot meet the requirements of prac- tical applications. To address this problem, we propose an effective approach called Layered Feature Repre- sentation for Differentiable Architecture Search (LFR- DARTS). Specifically, we iteratively search for multiple cells with different architectures from shallow to deep layers of the super network. For each iteration, we optimize the architecture of a cell by gradient descent and prune out weak connections from this cell. After obtain- ing the optimal architecture of this cell, we deepen the super network by increasing the number of this cell, so as to create an adaptive network context to search for a deeper-adaptive cell in the next iteration. Thus, our LFR-DARTS can discover the architecture of each cell at a specific and adaptive network depth, which embeds the ability of layered feature representations into each cell to sufficiently extract layered features in different depths. Extensive experiments show that our algorithm achieves an advanced performance on the datasets of CIFAR10, fashionMNIST and ImageNet while at low search costs.