AdaAugment: A Tuning-Free and Adaptive Approach to Enhance Data Augmentation (2405.11467v2)

Abstract: Data augmentation (DA) is widely employed to improve the generalization performance of deep models. However, most existing DA methods use augmentation operations with random magnitudes throughout training. While this fosters diversity, it can also inevitably introduce uncontrolled variability in augmented data, which may cause misalignment with the evolving training status of the target models. Both theoretical and empirical findings suggest that this misalignment increases the risks of underfitting and overfitting. To address these limitations, we propose AdaAugment, an innovative and tuning-free Adaptive Augmentation method that utilizes reinforcement learning to dynamically adjust augmentation magnitudes for individual training samples based on real-time feedback from the target network. Specifically, AdaAugment features a dual-model architecture consisting of a policy network and a target network, which are jointly optimized to effectively adapt augmentation magnitudes. The policy network optimizes the variability within the augmented data, while the target network utilizes the adaptively augmented samples for training. Extensive experiments across benchmark datasets and deep architectures demonstrate that AdaAugment consistently outperforms other state-of-the-art DA methods in effectiveness while maintaining remarkable efficiency.

AdaAugment: Enhancing Data Augmentation with Adaptive and Tuning-Free Methods

Introduction

Data Augmentation (DA) is a technique used in the training of deep neural networks to increase the diversity of the training data by creating modified versions of existing data samples. However, most existing DA methods use random augmentation magnitudes, which can introduce uncontrolled variability and may not align with the evolving training status of the model. This misalignment can lead to underfitting during the initial stages of training and overfitting in later stages. To address these limitations, this paper presents AdaAugment, a tuning-free and adaptive DA method that dynamically adjusts augmentation magnitudes based on real-time feedback from the target network using reinforcement learning.

How AdaAugment Works

Dual-Model Architecture

AdaAugment features a dual-model architecture consisting of a policy network and a target network. The policy network determines the magnitudes of augmentation operations, while the target network utilizes these adaptively augmented samples for training. Both networks are optimized jointly, making the adaptive adjustment process more integrated and efficient.

Key Components:

• Policy Network: Learns the policy determining augmentation magnitudes based on real-time feedback during training.
• Target Network: Uses the adaptively augmented samples for training, providing feedback to the policy network.

Reinforcement Learning Approach

The reinforcement learning (RL) component formulates the augmentation magnitude adjustment as a Markov Decision Process (MDP). Here's a simplified breakdown:

1. State Space (S): Considers the inherent difficulty of each sample, the current training status, and the intensity of augmentation.
2. Action Space (A): Contains actions representing different magnitudes of augmentation, ranging from 0 (no augmentation) to 1 (maximum augmentation).
3. Reward Function (R): Designed to balance underfitting and overfitting risks by leveraging losses from fully augmented, non-augmented, and adaptively augmented data.

Reward Function Formula:

$$r = \lambda(L_{\text{full}} - L_{\text{ada}}) + (1 - \lambda)(L_{\text{ada}} - L_{\text{none}})$$

where $L_{\text{full}}$ is the loss of fully augmented data, $L_{\text{none}}$ is the loss of non-augmented data, and $L_{\text{ada}}$ is the loss of adaptively augmented data.

Experimental Results

CIFAR-10 and CIFAR-100

Table 1: Test accuracy (%) on CIFAR-10/100

| Dataset | Method | ResNet-18 | ResNet-50 | WRN-28-10 | ShakeShake | |-||--|-|-|-| | CIFAR-10 | Baseline | 95.28 ±0.14* | 95.66±0.08* | 95.52 ±0.11*| 94.90 ±0.07*| | | CutMix | 96.64 ±0.62* | 96.81±0.10* | 96.93 ±0.10*| 96.47 ±0.07 | | | ... | ... | ... | ... | ... | | | AdaAugment | 96.75 ±0.06 | 97.34±0.13 | 97.66 ±0.07 | 97.41 ±0.06 |

AdaAugment consistently outperforms existing state-of-the-art DA methods across different network architectures. Noteworthy improvements include a 1.47% boost for ResNet-18 and a 2.14% for WRN-28-10 on CIFAR-10.

Tiny-ImageNet

Results on Tiny-ImageNet

| Method | ResNet-18 | ResNet-50 | WRN-50-2 | ResNext-50 | ||-|||-| | Baseline | 61.38±0.99 | 73.61±0.43 | 81.55±1.24 | 79.76±1.89 | | CutMix | 64.09±0.30 | 76.41±0.27 | 82.32±0.46 | 81.31±1.00 | | ... | ... | ... | ... | ... | | AdaAugment | 71.25±0.64 | 79.11±1.51 | 83.07±0.78 | 81.92±0.29 |

On Tiny-ImageNet, AdaAugment shows significant performance improvements, such as a 9.87% increase for ResNet-18 compared to the baseline.

Practical and Theoretical Implications

Theoretical Implications

AdaAugment introduces a paradigm shift by using adaptive magnitudes in DA, which aligns with the training status of models and mitigates risks of underfitting and overfitting. This approach can be extended to various tasks beyond image classification, such as NLP and time-series analysis.

Practical Implications

Practically, AdaAugment offers a more efficient way to implement DA without manual tuning. This can streamline the workflow for data scientists and reduce the need for extensive hyperparameter tuning. The minimal additional computational overhead (around 0.5 GPU hours) makes it feasible for real-world applications.

Future Developments

Future research could explore extending AdaAugment to other domains and tasks, further optimizing the policy network, and integrating additional types of data transformations.

Conclusion

AdaAugment offers a robust, adaptive, and tuning-free solution to enhance DA, demonstrating superior efficacy in improving model performance across various datasets and architectures. Its ability to dynamically adjust augmentation magnitudes makes it a valuable tool for achieving better generalization in deep learning models.