Renewing Iterative Self-Labeling Domain Adaptation With Application to the Spine Motion Prediction

Gecheng Chen, Yu Zhou, Xudong Zhang, Rui TuoIEEE Transactions on Automation Science and Engineering 2023 · DOI

TL;DR

This paper introduces Renewing Iterative Self-Labeling Domain Adaptation (Re-ISDA), a novel transfer learning algorithm designed to improve prediction accuracy when training and testing data have different distributions and limited labeled samples. Re-ISDA addresses the "mis-labeling issue" inherent in iterative self-labeling methods by actively renewing pseudo-labels, leading to more robust and accurate predictions, particularly for time-series data like cervical spine motion.

What Problem Does This Paper Address?

Traditional supervised machine learning models struggle when the distribution of input features in the training data differs significantly from that of the testing data. This is a common problem in real-world scenarios, especially in fields like biomechanics where individual variations (e.g., BMI, sex) can lead to distinct feature distributions. Furthermore, acquiring labeled data can be expensive and time-consuming, resulting in limited training samples. The paper specifically targets the challenge of predicting cervical spine motion using surface-based measurements, where subject-specific variations in features like body mass index (BMI) and sex lead to differing input distributions, and the cost of obtaining accurate skeletal motion data (dynamic radiography) limits the available labeled samples. Existing transfer learning methods, while helpful, often fall short in these scenarios, particularly with limited data or when dealing with the potential for errors in self-generated labels.

Methods

The core contribution is the Renewing Iterative Self-Labeling Domain Adaptation (Re-ISDA) algorithm. It builds upon the Iterative Self-Labeling Domain Adaptation (ISDA) framework but introduces a crucial modification: a "renewing" step.

Here's a breakdown:

  • Domain Adaptation: The problem is framed as domain adaptation, where the training data (source domain) and testing data (target domain) have different input distributions ($P(X_s) \neq P(X_t)$) but share the same underlying relationship between inputs and outputs ($P(Y|X_s) = P(Y|X_t)$).
  • Iterative Self-Labeling (ISDA): The ISDA method iteratively learns a model from labeled source data and pseudo-labeled target data. In each iteration, it trains a model, uses it to predict labels for some target samples (pseudo-labeling), and then incorporates these pseudo-labeled samples into the training set for the next iteration.
  • The "Mis-labeling Issue": A key problem with ISDA is that initial mis-predictions (pseudo-labels) can propagate and negatively impact subsequent learning, leading to a "mis-labeling issue."
  • Re-ISDA's Innovation: The Renewing Step: Re-ISDA addresses this by adding a step in each iteration to renew the pseudo-labels of all previously labeled target samples using the most up-to-date predictor. This self-correction mechanism helps to mitigate the impact of early errors and stabilizes the learning process.
  • Dynamic Programming Formulation: The authors also formulate the problem as a dynamic programming model, which is then solved using a greedy algorithm. This provides a theoretical grounding for their approach.

Datasets

  • Friedman Function Dataset: A synthetic dataset generated using the Friedman function (a standard benchmark for regression problems) to evaluate the general performance of Re-ISDA against other domain adaptation methods. This dataset consists of 80 training samples and 41 testing samples.
  • Cervical Spine Motion Dataset: This is the primary application dataset. It comprises labeled data from six subjects for training and one subject for testing. The data includes marker coordinates, BMI, neck angle, head angle, and C7-MMP statistics, used to predict the Cobb angle (a measure of cervical spinal curvature). This dataset is characterized by subject-to-subject variations and limited labeled samples.

Key Insights and Ideas

  • The Power of "Renewing" Labels: The most significant insight is that actively correcting and updating pseudo-labels throughout the iterative process is crucial for robust domain adaptation, especially when dealing with noisy or uncertain initial predictions. This "renewing" step directly combats the detrimental effects of the mis-labeling issue.
  • Dynamic Programming for Domain Adaptation: Framing domain adaptation as a dynamic programming problem offers a structured way to think about the sequential nature of iterative learning and provides a theoretical basis for developing greedy algorithms.
  • Suitability for Time-Series Data: The iterative and self-correcting nature of Re-ISDA makes it particularly well-suited for time-series data, where the temporal dependencies and the evolving nature of predictions can be effectively managed.
  • Balancing Precision and Cost: The paper highlights that the size of the batch of samples added in each iteration ($\eta$) is a critical hyperparameter. Smaller $\eta$ leads to higher precision but increased computational cost, necessitating a trade-off.

Key Results

  • Friedman Function: Re-ISDA achieved the lowest Root Mean Squared Error (RMSE) on the synthetic Friedman function dataset, outperforming standard Neural Networks (NN) without domain adaptation, Transfer Component Analysis (TCA), and Iterative Self-Labeling Domain Adaptation (ISDA). Specifically, Re-ISDA had an RMSE of 2.033, compared to ISDA's 4.070 and NN's 3.623.
  • Cervical Spine Motion Prediction: Re-ISDA demonstrated superior performance in predicting cervical spine motion. It achieved the lowest RMSE (2.68) among the tested methods, significantly outperforming NN without DA (8.19), TCA (28.88), ISDA (6.21), and KMM (13.2). The paper notes that Re-ISDA's predictions closely matched the true values, unlike other methods that showed significant deviations or failed to capture the trend.
  • Improved Stability: The "renewing" step in Re-ISDA led to more stable predictions and a lower overall error compared to standard ISDA, which suffered from performance degradation in later stages due to accumulated mis-labeling.

Limitations and Future Work

  • Computational Cost: The paper acknowledges that Re-ISDA has a slightly higher computational cost than some other domain adaptation methods due to the additional label renewal step. This makes it more suitable for moderate-sized datasets.
  • Hyperparameter Tuning: The performance of Re-ISDA is sensitive to hyperparameters like the batch size ($\eta$) for adding new samples. The paper suggests that smaller $\eta$ generally leads to better precision but requires more training epochs and higher computational cost, necessitating a trade-off.
  • Applicability to Classification: While the current work focuses on regression problems, the authors suggest that Re-ISDA, with proper modifications, could also be applicable to classification problems.
  • Time-Series Data Suitability: The paper emphasizes Re-ISDA's suitability for time-series data, hinting at future work exploring its application in other dynamic or sequential prediction tasks.

Summary updated

Abstract

The area of transfer learning comprises supervised machine learning methods that cope with the issue when the training and testing data have different input feature spaces or distributions. In this work, we propose a novel transfer learning algorithm called Renewing Iterative Self-labeling Domain Adaptation (Re-ISDA). This work is motivated by a cervical spine motion prediction problem, with the goal of predicting the cervical spine motion of different subjects using the measurements of “exterior features”. The joint distribution of the exterior features for each subject may vary with one’s distinct BMI, sex and other characteristics; the sample size of this problem is limited due to the high experimental cost. The proposed method is well suited for transfer learning problems with limited training samples. The learning problem is formulated as a dynamic programming model and the latter is then solved by an efficient greedy algorithm. Numerical studies show that the proposed method outperforms prevailing transfer learning methods. The proposed method also achieves high prediction accuracy for the cervical spine motion problem. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Note to Practitioners</i> —This paper is motivated by the problem of using human body exterior features (e.g., contours, surface markers) acquired by computer vision or other measurement means to determine the underlying physics or mechanics. The goal is to build a subject-specific model to predict cervical spine motion based on surface-based measurements. Conventional supervised learning and transfer learning methods fail to give an accurate prediction due to the subject-to-subject variation and/or the small sample size of the training data. This paper proposes a novel transfer learning method called renewing iterative self-labeling domain adaptation to modify the possible error of the pseudo labels and thus make the training process more stable. The proposed method improves the accuracy by over <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$35\%$</tex-math> </inline-formula> on the traditional transfer learning methods. This method is applicable to limited-sample problems with different training and testing input distributions, and is particularly suitable for time-series data.
BibTeX
@article{chen2023renewing,
  author = {Gecheng Chen and Yu Zhou and Xudong Zhang and Rui Tuo},
  title = {{Renewing Iterative Self-Labeling Domain Adaptation With Application to the Spine Motion Prediction}},
  journal = {IEEE Transactions on Automation Science and Engineering},
  year = {2023},
  doi = {10.1109/tase.2023.3280900},
}
Figure