Differentiable Architecture Search Meets Network Pruning at Initialization: A More Reliable, Efficient, and Flexible Framework

Although Differentiable ARchiTecture Search (DARTS) has become the mainstream paradigm in Neural Architecture Search (NAS) due to its simplicity and efficiency, more recent works found that the performance of the searched architecture barely increases with the optimization proceeding in DARTS, and the final magnitudes obtained by DARTS could hardly indicate the importance of operations. The above observation reveal that the supervision signal in DARTS may be a poor or unreliable indicator for the architecture search, inspiring an interesting and promising direction: can we measure the operation importance without any training under the differentiable paradigm? We provide an affirmative answer by customizing the NAS as a network pruning at initialization problem. With leveraging recently-proposed synaptic saliency criteria in the network pruning at initialization, we seek to score the importance of candidate operations in differentiable NAS without any training, and proposed a novel framework called training free differentiable architecture search (FreeDARTS) accordingly. We show that, without any training, FreeDARTS with different proxy metrics can outperform most NAS baselines in different search spaces. More importantly, FreeDARTS is extremely memory-efficient and computational-efficient as it abandons the training in the architecture search phase, enabling FreeDARTS to perform architecture search on a more flexible space and eliminate the depth gap between architecture search and evaluation. We hope our work inspires more attempts in solving NAS from the perspective of pruning at initialization.