$Λ$-DARTS: Mitigating Performance Collapse by Harmonizing Operation Selection among Cells
This addresses a critical problem in neural architecture search for researchers and practitioners by mitigating performance collapse, though it is incremental as it builds on existing DARTS methods.
The paper tackles the performance collapse issue in Differentiable Neural Architecture Search (DARTS), where discovered architectures decline in quality, by identifying a structural flaw in its weight-sharing framework that biases layers closer to the output. It proposes $\Lambda$-DARTS with two regularization terms to harmonize operation selection, preventing collapse and showing effectiveness across six search spaces and three datasets.
Differentiable neural architecture search (DARTS) is a popular method for neural architecture search (NAS), which performs cell-search and utilizes continuous relaxation to improve the search efficiency via gradient-based optimization. The main shortcoming of DARTS is performance collapse, where the discovered architecture suffers from a pattern of declining quality during search. Performance collapse has become an important topic of research, with many methods trying to solve the issue through either regularization or fundamental changes to DARTS. However, the weight-sharing framework used for cell-search in DARTS and the convergence of architecture parameters has not been analyzed yet. In this paper, we provide a thorough and novel theoretical and empirical analysis on DARTS and its point of convergence. We show that DARTS suffers from a specific structural flaw due to its weight-sharing framework that limits the convergence of DARTS to saturation points of the softmax function. This point of convergence gives an unfair advantage to layers closer to the output in choosing the optimal architecture, causing performance collapse. We then propose two new regularization terms that aim to prevent performance collapse by harmonizing operation selection via aligning gradients of layers. Experimental results on six different search spaces and three different datasets show that our method ($Λ$-DARTS) does indeed prevent performance collapse, providing justification for our theoretical analysis and the proposed remedy.