Core Concepts
A novel graph-based performance predictor that leverages both forward and reverse representations of neural architectures to enhance prediction accuracy, especially in data-limited settings.
Abstract
The paper introduces a novel graph-based performance predictor called FR-NAS that utilizes both forward and reverse representations of neural architectures to enhance prediction accuracy. The key insights are:
Analyzing the feature embeddings from GNN predictors using only forward graph representations reveals that in the presence of limited training data, the encoder faces challenges in effectively capturing crucial features for precise predictions.
To address this, the authors propose a predictor that employs two separate GIN encoders to process the forward and reverse graph representations of neural architectures.
To ensure the two encoders converge towards shared features, a customized training loss is introduced that minimizes the discrepancy between the embeddings from the two encoders.
Comprehensive experiments on benchmark datasets including NAS-Bench-101, NAS-Bench-201, and DARTS search space demonstrate that the proposed FR-NAS outperforms state-of-the-art GNN-based predictors, especially with smaller training datasets, achieving 3%-16% higher Kendall-tau correlation.
Ablation studies further confirm the effectiveness of the dual graph representations and the tailored training loss in improving the predictor's performance.
Stats
The NAS-Bench-101 search space consists of around 423k unique convolutional architectures.
The NAS-Bench-201 search space has a larger size compared to NAS-Bench-101.
The DARTS search space comprises approximately 10^21 architectures.
Quotes
"By contrast, the neural architectures are inherently bidirectional, involving both forward and backward propagation phases. This raises the question: Can we harness the inherent bidirectionality of neural architectures to enhance the performance of graph predictors?"
"Our observations suggest that in the presence of limited training data, the encoder often faces challenges in effectively embedding features crucial for precise predictions."