DARTS
A representative method in neural architecture search that combines weight sharing with gradient-based optimization. DARTS reduces search time from thousands of GPU-hours to a few days on a single GPU, making architecture search substantially more practical.
Timeline (last 90 days)
DARTS (Differentiable Architecture Search), based on weight sharing, compresses search time to a few days on a single GPU.
Early NAS work (such as NASNet and DARTS) was found, due to opaque search space design, to have discovered architectures that largely depended on specific proxy tasks and search budgets.
All Facts (2)
DARTS (Differentiable Architecture Search), based on weight sharing, compresses search time to a few days on a single GPU.
50%UnverifiedEarly NAS work (such as NASNet and DARTS) was found, due to opaque search space design, to have discovered architectures that largely depended on specific proxy tasks and search budgets.
50%