Presentation
PertNAS: Architectural Perturbations for Memory-Efficient Neural Architecture Search
DescriptionDifferentiable Neural Architecture Search (NAS) relies on aggressive weight-sharing to reduce its search cost. This leads to GPU-memory bottlenecks that hamper the algorithm's scalability. To solve this, we propose a perturbation-based evolutionary approach that reduces memory cost while largely maintaining the efficiency benefits of weight-sharing. Our approach makes minute changes to compact neural architectures and measures the impact on performance. In this way, it extracts high-quality motifs from the search space. Our results show competitive performance on multiple benchmarks, including CIFAR10, ImageNet2012, and NASBench-301.
Event Type
Research Manuscript
TimeTuesday, July 11th4:25pm - 4:40pm PDT
Location3004, 3rd Floor
AI
AI/ML Algorithms