💻 computer science

Few-Shot Transfer Learning for Gait Cycle Percentage Estimation by Leveraging Classification Knowledge for Regression Tasks

This paper demonstrates that leveraging automatically derived binary gait phase labels through classification-to-regression transfer learning significantly improves few-shot gait cycle percentage estimation from wearable IMU data, achieving up to 40% MAE reduction over zero-shot baselines while providing practical engineering guidelines for optimal implementation.

Original authors: Hangkai Liu, Peng Wu, Jiaxiang Sun, Ziming Wang, Ziyun Ding, Xunju Ma, Chuncan Ma, Rui Song, Yibin Li, Huanghe Zhang

Published 2026-10-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hangkai Liu, Peng Wu, Jiaxiang Sun, Ziming Wang, Ziyun Ding, Xunju Ma, Chuncan Ma, Rui Song, Yibin Li, Huanghe Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Walking is a rhythm we rarely think about, yet it is a complex sequence of movements that doctors and engineers study closely to understand health, prevent falls, and design better robotic limbs. To analyze this rhythm, scientists often break a single step into two main parts: the stance phase, when the foot is firmly on the ground, and the swing phase, when the foot lifts and moves forward. Traditionally, researchers have used expensive, high-tech motion capture studios to measure exactly where a person is in this cycle at any given moment, a process that is precise but costly and difficult to scale. However, a simpler method exists to determine just which of the two phases a person is in, using basic sensors that can be attached to a shoe. This creates a fundamental imbalance: we can easily generate vast amounts of data telling us if a foot is up or down, but we struggle to gather the continuous, detailed data needed to know exactly how far along the step a person is.

A team of researchers from Shandong University and the University of Birmingham set out to solve this problem by teaching a computer to use the easy-to-get information to learn the hard-to-get information. They asked whether a computer model, trained extensively on the simple "up or down" labels, could be adapted to predict the precise percentage of a walking cycle with very little new data. Their work demonstrates that this is not only possible but highly effective. By first teaching a deep learning model to distinguish between the two phases of walking, and then gently adjusting that same model to predict the continuous percentage of the step, they achieved results that would otherwise require far more expensive data collection.

The researchers tested their method using data from 178 people across two public datasets, one containing elderly adults and the other healthy adults. They simulated a scenario where a clinic might have access to thousands of hours of simple "stance versus swing" recordings but only a tiny amount of precise, motion-capture data. In the most extreme test, they gave the model data from just one person to learn the precise percentage. Without any prior training, a model starting from scratch performed poorly, but the model that had first learned the simple "up or down" task was able to adapt quickly. With data from only a single subject, this transfer learning approach reduced the error in its predictions by nearly 24 percent compared to a model trained from scratch. Even more strikingly, the model was able to improve its accuracy by 40 percent compared to a baseline that had no fine-tuning at all, proving that the knowledge gained from the simple task was directly useful for the complex one.

The study also investigated exactly how much data is needed to make this work best. The researchers found that the sweet spot for this method lies between 50 and 500 samples of precise data. In this range, the model gains a significant advantage, improving its accuracy by 10 to 14 percent over traditional methods. This suggests that a medical clinic or research team does not need to collect massive amounts of expensive motion-capture data to build a reliable system. Instead, they can gather a large volume of simple sensor data to teach the system the basics of walking, and then spend their limited resources on a small, carefully selected set of precise measurements to fine-tune the model.

To ensure this method works reliably, the team examined the technical details of how the computer learns. They discovered that the way the model handles its internal statistics and how fast it learns new information are critical. Specifically, allowing the model to slightly update its internal settings while keeping the core knowledge from the first training phase intact proved essential. They also found that using different learning speeds for the part of the model that recognizes patterns versus the part that makes the final prediction helped the system adapt without forgetting what it already knew. These findings provide a clear roadmap for engineers and clinicians who want to deploy accurate gait analysis systems without the prohibitive cost of large-scale motion capture studies.

The implications of this work extend to the development of wearable technology and rehabilitation tools. By showing that a model can learn to predict the continuous progress of a step by first mastering the binary distinction of foot position, the researchers have offered a practical path forward for personalized gait monitoring. This approach allows for the creation of sophisticated analysis tools using data that is much easier and cheaper to collect. While the current study focused on healthy and elderly populations walking normally, the success of this method suggests a future where similar techniques could be applied to more complex conditions, potentially helping to monitor and treat movement disorders with greater precision and accessibility.

Technical Summary: Few-Shot Transfer Learning for Gait Cycle Percentage Estimation

Problem Statement
Accurate estimation of gait cycle percentage (a continuous regression task from 0% to 100%) using wearable Inertial Measurement Unit (IMU) sensors typically requires large volumes of precisely labeled training data. Obtaining these regression labels is expensive and time-consuming, as it relies on sophisticated motion capture systems or manual annotation. In contrast, binary gait phase labels (stance vs. swing) can be derived automatically and at negligible cost using simple kinematic thresholds on sensor data. This creates a fundamental data asymmetry: classification labels are abundant, while regression labels are scarce. This paper addresses the challenge of achieving accurate regression-based gait phase estimation in few-shot regimes (limited regression-labeled data) by leveraging the abundant classification data.

Methodology
The authors propose a three-stage classification-to-regression transfer learning framework:

Source Pre-training: A deep neural network (a fully-connected architecture with four layers, batch normalization, and dropout) is pre-trained on the abundant binary classification dataset (stance/swing). The classification labels are generated by binarizing the regression targets at a 0.5 threshold to ensure perfect task alignment.
Head Swap: The pre-trained classification head is replaced with a randomly initialized regression head, while the feature extractor (backbone) is preserved.
Few-Shot Fine-tuning: The model is fine-tuned on a small set of regression-labeled data. The authors investigate critical engineering configurations, including:
- Learning Rates: Using discriminative learning rates (lower rate for the backbone, higher for the head) to protect pre-trained features while allowing the new head to adapt.
- Batch Normalization (BN): Comparing strategies such as freezing, resetting, or allowing BN statistics to update.
- Data Standardization: Ensuring input features are standardized using StandardScaler fitted on the training data.

Key Contributions

Empirical Validation of Cross-Task Transfer: The study demonstrates that knowledge learned from binary gait classification can be effectively transferred to continuous gait cycle percentage regression, significantly reducing the need for expensive regression-labeled data.
Performance Gains in Few-Shot Regimes: Experiments on 178 subjects across two public datasets (GSTRIDE and Mobilise-D HA) show that transfer learning achieves up to a 24% improvement in Mean Absolute Error (MAE) compared to training from scratch when using data from only one subject. It also reduces MAE by up to 40% compared to a zero-shot baseline (pre-trained backbone with no fine-tuning).
Optimal Transfer Range Identification: Fine-grained analysis reveals an optimal sample size range of 50–500 samples, where transfer learning yields consistent improvements of 10–14% over from-scratch training.
Engineering Guidelines: The paper provides actionable recommendations for practitioners, including:
- Using discriminative learning rates for extreme low-data scenarios (<10% data).
- Allowing Batch Normalization layers to update during fine-tuning (rather than freezing) when combined with proper input standardization.
- Utilizing at least 50% of available source data for pre-training to capture transferable features.

Results

GSTRIDE Dataset (Elderly): With data from a single subject (~27k samples), transfer learning achieved an MAE of 0.143 ( $R^2$ = 0.527), compared to 0.187 MAE ( $R^2$ = 0.104) for from-scratch training. This represents a 5-fold improvement in explained variance.
Data Efficiency: Transfer learning with data from 5 subjects achieved an $R^2$ of 0.614, outperforming from-scratch training with 20 subjects ( $R^2$ = 0.598), effectively reducing the required labeled subjects by a factor of 4.
Generalization: The approach demonstrated robust generalization across different populations (elderly vs. healthy adults) and datasets.
Feature Analysis: t-SNE visualization and Centered Kernel Alignment (CKA) analysis confirmed that pre-training learns discriminative features for gait phases (stance/swing clustering) which are then smoothly adapted to continuous regression targets, with early layers remaining highly similar and later layers adapting to the specific task.

Significance and Claims
The paper claims to establish a cost-effective paradigm for sensor-based regression tasks where classification labels are abundant but regression annotations are expensive. By leveraging the shared feature representations between gait classification and regression, the proposed method enables the deployment of accurate gait monitoring systems in real-world scenarios (e.g., clinical rehabilitation) without the prohibitive cost of large-scale motion capture data collection. The authors emphasize that their findings provide a practical protocol: collect a large corpus of threshold-derived classification labels for pre-training, and then fine-tune with a small fraction of expensive regression-labeled data to achieve high accuracy.

The study acknowledges limitations, noting that validation was performed on free-walking conditions for elderly and healthy populations, and further investigation is needed for pathological gait patterns (e.g., Parkinson's, stroke) and more complex biomechanical variables (e.g., joint angles, ground reaction forces).

Technical Summary: Few-Shot Transfer Learning for Gait Cycle Percentage Estimation

More like this