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

2024-09-18

Abstract: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.

Computer Vision and Pattern Recognition,Cryptography and Security

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to improve the performance of Eye Movement Recognition (EMR), especially in terms of network architecture design. Traditional eye movement recognition methods rely on manually - designed network architectures and prior knowledge, which limit the flexibility and performance of the model. To solve these problems, the author introduced the Neural Architecture Search (NAS) algorithm and proposed an improved differentiable architecture search algorithm - Relax DARTS.

### Specific Problems and Solutions:

1. **Limitations of Manually - Designed Network Architectures**:
- **Problem**: Traditional deep - learning methods rely on manual design and prior knowledge in network architecture design, which is not only time - consuming but also difficult to find the optimal structure.
- **Solution**: Introduce the Neural Architecture Search (NAS) algorithm, especially Relax DARTS, to automatically search for the optimal network architecture.

2. **Weight - Sharing Problem in DARTS**:
- **Problem**: In DARTS, weight - sharing between different units may lead to over - fitting and performance degradation.
- **Solution**: Relax DARTS avoids the weight - sharing problem by independently training the architecture parameter \(\alpha\) of each unit, thereby achieving a more accurate target architecture.

3. **Lack of Flexibility in Input Selection**:
- **Problem**: In traditional methods, the unit's selection of input is not flexible enough, which is prone to cause over - fitting.
- **Solution**: Introduce the module input weight \(\beta\), enabling the unit to flexibly select input, thereby alleviating the over - fitting phenomenon and improving the model performance.

4. **Lack of Global Search Ability**:
- **Problem**: Existing methods lack global adjustment ability during the search process, resulting in limited final model performance.
- **Solution**: Relax DARTS endows each unit with the ability to independently update architecture parameters and realizes global architecture adjustment through learning the parameter \(\beta\).

### Paper Contributions:

- **Applying NAS to Eye Movement Recognition for the First Time**: Proposed a differentiable architecture search method based on independent cell - structure search, achieving more efficient network search and training.
- **Increasing Network Flexibility**: Abandoned the parameter - sharing and unit - stacking strategies and endowed each unit with the ability to independently update architecture parameters.
- **Global Search Ability**: By introducing the learnable input weight \(\beta\), the network can perform global architecture adjustment while conducting local unit search.
- **Experimental Verification**: The experimental results on four public datasets show that Relax DARTS is superior to existing methods in reducing the validation error rate and has reached a new state - of - the - art level.

### Summary:

Relax DARTS aims to solve the limitations of network architecture design in eye movement recognition by improving the DARTS algorithm and improve the flexibility and performance of the model. By independently training the architecture parameter \(\alpha\) and introducing the input weight \(\beta\), Relax DARTS not only enhances the generalization ability of the model but also significantly improves the accuracy and robustness of eye movement recognition.