FairNAS: Rethinking Evaluation Fairness of Weight Sharing Neural Architecture Search

By Xiangxiang Chu, Bo Zhang, Ruijun Xu, Jixiang Li. 2019

370 citations recorded by the source corpus.

One of the most critical problems in weight-sharing neural architecture search is the evaluation of candidate models within a predefined search space. In practice, a one-shot supernet is trained to serve as an evaluator. A faithful ranking certainly leads to more accurate searching results. However, current methods are prone to making misjudgments. In this paper, we prove that their biased evaluation is due to inherent unfairness in the supernet training. In view of this, we propose two levels of constraints: expectation fairness and strict fairness. Particularly, strict fairness ensures equal optimization opportunities for all choice blocks throughout the training, which neither overestimates nor underestimates their capacity. We demonstrate that this is crucial for improving the confidence of models’ ranking. Incorporating the one-shot supernet trained under the proposed fairness constraints with a multi-objective evolutionary search algorithm, we obtain various state-of-the-art models, e.g., FairNAS-A attains 77.5% top-1 validation accuracy on ImageNet.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Neural Architecture Search & AutoML, AI Ethics, Fairness & Society

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Xiaodan Liang
2. Xiaojun Chang
3. Zhenguo Li
4. Hang Xu
5. Xinggang Wang
6. Dacheng Tao
7. Dahua Lin
8. Liang Lin

In Computer Vision and Pattern Recognition

1. Luc Van Gool
2. Dacheng Tao
3. Xiaogang Wang
4. Ming-Hsuan Yang
5. Yi Yang
6. Xiaodan Liang
7. Chunhua Shen
8. Wanli Ouyang

In Neural Architecture Search & AutoML

1. Xiaodan Liang
2. Zhenguo Li
3. Xiaojun Chang
4. Tong Zhang
5. Hang Xu
6. Bo Zhang
7. Houwen Peng
8. Yong Guo

In AI Ethics, Fairness & Society

1. Houwen Peng
2. Xiaojun Chang
3. Xiaodan Liang

Citations in the corpus

• A Comprehensive Survey of Neural Architecture Search (2020, background): “In addition, the research of Fairnas [100] and [101] also shows that candidate network architectures based on these parameter sharing often cannot be adequately trained, which will lead to inaccurate ranking of candidate network architectures.”
• GreedyNAS: Towards Fast One-Shot NAS With Greedy Supernet (2020, methodology): “Instead of specifying an epoch number [11, 4], we use the proposed principle to stop the supernet training with tolerance α = 0.”
• Hierarchical Neural Architecture Search for Single Image Super-Resolution (2020, background): “network and thus alleviate the unfairness issue [28] that some sub-networks (or candidate operations) may be over-optimized while the others are under-optimized.”
• Real-Time Federated Evolutionary Neural Architecture Search (2020, background): “Apart from this, the trainable parameters in the batch normalization layer may slow down the convergence speed in federated learning, since they perform poorly for learning with small batch sizes [53] and weight sharing training paradigms [39], [54], [55], especially for non-IID scenarios.”
• DSNAS: Direct Neural Architecture Search Without Parameter Retraining (2020, methodology): “The correlation between the performance at the end of searching and after the retraining in evaluation is fairly low, as long as the parameter-sharing technique is used [20, 5].”
• BossNAS: Exploring Hybrid CNN-transformers with Block-wisely Self-supervised Neural Architecture Search (2021, methodology): “By contrast, network-level search spaces with layer-level granularity [7, 70, 15, 37, 46, 87] and block-level granularity [63, 29, 64] search for the macro network-level structure using manually designed building blocks (e.”
• Densely Connected Search Space for More Flexible Neural Architecture Search (2019, background): “Moreover, we find that recent works [37, 3, 41, 7] manually tune width settings to obtain better performance, which indicates the design of network width demands much prior-based knowledge and trial-and-error.”
• Bridging the Gap between Sample-based and One-shot Neural Architecture Search with BONAS (2020, background): “This makes it infeasible to train each architecture fairly [5], and some architectures may not be trained as sufficiently as others.”
• Block-Wisely Supervised Neural Architecture Search With Knowledge Distillation (2019, methodology): “One-shot approaches [18, 15, 9, 8] ensure fairness among all sub-models.”
• EfficientViT: Memory Efficient Vision Transformer with Cascaded Group Attention (2023, methodology): “We compare EfficientViT-M4 with efficient models [12, 22, 26, 63, 66, 68] on the COCO [42] object detection task, and present the results in Tab.”
• Cream of the Crop: Distilling Prioritized Paths For One-Shot Neural Architecture Search (2020, methodology): “The same as [27], we train the detection model by 12 epochs.”
• AutoFormer: Searching Transformers for Visual Recognition (2021, methodology): “Such strategy performs well when used to search architectures over convolutional neural networks space [16, 6, 8, 21, 49].”