Source-linked AI summary
Relax DARTS: Relaxing the Constraints of Differentiable Architecture Search for Eye Movement Recognition
Hongyu Zhu, Xin Jin, Hongchao Liao, Yan Xiang, Mounim A. El-Yacoubi, Huafeng Qin
TL;DR
Eye-movement recognition has relied on manually designed network architectures, motivating automated architecture search. Relax DARTS independently trains cell architecture parameters and adds learnable input weights; it is reported to outperform existing methods and achieve state-of-the-art verification accuracy.
Problem
Deep-learning eye-movement recognition methods rely on manual network-architecture design and prior knowledge.
Source: §1 · p. 2Method
Relax DARTS independently searches each cell’s architecture parameters α and learns input weights β to adjust cell inputs.
Source: §1 · p. 2Results
Relax DARTS outperforms existing methods and achieves a new state-of-the-art verification accuracy.
Source: §5 · p. 8Takeaways & Limitations
Relax DARTS provides end-to-end model search and training for eye-movement biometric authentication.
Source: §5 · p. 8
Abstract
from arXiv · showhide
Eye movement biometrics is a secure and innovative identification method. Deep learning methods have shown good performance, but their network architecture relies on manual design and combined priori knowledge. To address these issues, we introduce automated network search (NAS) algorithms to the field of eye movement recognition and present Relax DARTS, which is an improvement of the Differentiable Architecture Search (DARTS) to realize more efficient network search and training. The key idea is to circumvent the issue of weight sharing by independently training the architecture parameters $\alpha$ to achieve a more precise target architecture. Moreover, the introduction of module input weights $\beta$ allows cells the flexibility to select inputs, to alleviate the overfitting phenomenon and improve the model performance. Results on four public databases demonstrate that the Relax DARTS achieves state-of-the-art recognition performance. Notably, Relax DARTS exhibits adaptability to other multi-feature temporal classification tasks.
1 Introduction
Eye movement biometrics is presented as a distinctive, stable behavioral modality, and this work introduces NAS to automatically search recognition networks. Relax DARTS gives cells independent architecture parameters and learnable input weights, and is reported to reduce verification error on four public datasets.
- Eye movement biometrics supports liveness detection and spoof-resistant continuous authentication, and can be combined with other ocular attributes for multimodal recognition.Source: §1 · p. 1
- This work makes the first attempt to introduce NAS to eye movement recognition and uses differentiable search to find networks automatically.Source: §1 · p. 2
- Relax DARTS gives each cell independent architecture parameters α, replacing parameter sharing and cell stacking so cells can update structure and choose operations independently.Source: §1 · p. 2
- Learned input weights β let cells adjust the proportions of direct, skip, or absent inputs, enabling global architectural adjustment during local cell search.Source: §1 · p. 2
- Relax DARTS outperforms existing works in reducing verification error across four public datasets.Source: §1 · p. 3
2 Related Work
Prior work spans manually engineered eye-movement features, deep-learning recognition models, and differentiable architecture search. The related-work section reviews representative methods in each area and describes DARTS variants addressing skip-connection enrichment.
2.1 Traditional eye movement recognition method
- Traditional eye-movement recognition manually filters features such as velocity, gaze duration, and path length before applying machine-learning algorithms.Earlier work also explored identity information in eye movements, oculomotor models, pupil size, and Gabor-wavelet features.Source: §2.1 · p. 3
2.2 Deep Learning based eye movement recognition method
- Deep-learning approaches use end-to-end feature learning, including Siamese and CNN models, dilated CNNs, LSTM-dense networks, and LSTM-Transformer combinations.Source: §2.2 · p. 3
2.3 Differentiable architecture search
- DARTS makes candidate-operation selection differentiable through continuous architecture weights; DARTS+, Fair DARTS, and DARTS- address skip-connection enrichment through distinct strategies.DARTS+ uses early stopping, Fair DARTS replaces Softmax with Sigmoid scoring, and DARTS- adds auxiliary skip connections.Source: §2.3 · pp. 3–4
3 Approach
Relax DARTS searches cells locally with independent architecture parameters and adjusts the network globally through learnable cell-input weights. Its search jointly optimizes architecture and network weights, then restarts training on the searched structure.
3.1 Review Relax DARTS
- DARTS relaxes operation selection into continuous architecture weights α, learned by optimizing the supernet’s training and validation losses.After finding the cell structure, training on the target task is restarted.Source: §3.1 · pp. 4–5
3.2 Local Search Strategy
- Each cell independently learns α, producing six distinct cells in a six-layer network formed by alternating Normal and Reduction Cells three times.The paper introduces independent α updates to reduce performance loss associated with parameter sharing and stacking identical cells, and uses the alternating six-layer structure to reduce overfitting risk.Source: §3.2 · p. 5
3.3 Global Search Strategy
- Learnable β weights each cell’s two inputs to fine-tune the global network structure during cell search.The inputs represent skip-connect and direct-connect paths; after softmax normalization, weights below c = 0.2 are replaced by all-zero tensors.Source: §3.3 · p. 6
4 Experiments
Experiments compare Relax DARTS with eye-movement recognition methods and DARTS variants on four public datasets. The reported results show lower verification error than existing approaches, while ablations examine independent α optimization and β input weights.
4 Experiments
- Evaluation uses JuDo1000 and three GazeBase sub-datasets, comparing Relax DARTS with recognition methods and DARTS, Fair-DARTS, and DARTS-.The study partitions data into rounds: the first round supplies training and validation for search, and the second round is used for testing.Source: §4 · p. 7, §4.1 · p. 7
- The evaluation trains each network directly on the dataset being evaluated rather than using a proxy dataset.The authors state that computational overhead is affordable and direct evaluation reduces the search-to-evaluation performance gap.Source: §4 · p. 7
4.1 Identification Performance of Relax DARTS
- 0.0657, 0.0610, and 0.0529 on the GazeBase subsets, and 0.0437 on JuDo1000 are the lowest reported verification errors among compared approaches.The comparisons cover RAN, HSS, TEX, and JuDo1000; the paper also reports higher recognition accuracy at different FARs.Source: Table 1 · p. 7, Table 2 · p. 7, §4.1 · p. 7
4.2 Ablation Experiment
- Independent α optimization improves the DARTS baseline, and adding β produces the lowest EER in the RAN ablation.The authors attribute α’s benefit to greater search freedom and position-adapted unique cells; β permits global architecture fine-tuning during cell search.Source: §4.2 · p. 8, Table 5 · p. 8
5 Conclusion
Relax DARTS combines alternating local and global architecture search for eye-movement biometric authentication. The authors report that it outperforms existing methods and achieves state-of-the-art verification accuracy.
5 Conclusion
- Relax DARTS implements alternating local and global architecture search for end-to-end eye-movement biometric authentication and achieves state-of-the-art verification accuracy.The paper reports that its approach outperforms existing methods.Source: §5 · p. 8