Transfer Learning for Robust Steering of Myoelectric Devices

 

Abstract

Although myoelectric control of several degrees of freedom in upper-limb prostheses works well under laboratory conditions, the control breaks down under real world conditions. The goal of this project is to bridge this gap by making the trained machine learning models more robust to real life disturbances such as electrode shift, posture changes and the effects of fatigue and sweat. The methodology under investigation strives at a particularly relevant and widespread open research area in machine learning: transfer learning and online learning in the presence of drift. Besides its mathematical and technological challenges in machine learning, these topics have a high relevance for many interdisciplinary problems, including model personalization, co-adaptive systems or life-long learning.

Research Questions and Methods

The core objective of this proposal is to design robust and personalized adaptive control for myoelectric devices which can react to novel situations and is capable of life-long model adaptation. To achieve this overarching goal, the following concrete challenges will be addressed:
  1. Efficient Transfer Learning from few data samples: Metric learning based methods have shown promising results in first studies [Prahm16, Paaßen17]. We aim to further incorporate a similar transfer learning approach in rich dynamical models, which are better suited to deal with the time dynamics present in such signals.
  2. Online learning in the absence of supervised information: While this is generally ill-posed, we plan to identify system invariances which characterize the interplay of the model and the sensor data, such as a high data likelihood of the observed data by a generative model. Such invariances can take the role of a teacher signal in the case of fully unsupervised model adaptation.
  3. Robustness guarantees of machine learning models: Predictable model behavior constitutes a pivotal property for user acceptance. However, distinguishing between unexpected behavior and predictable concept drift constitutes an open problem. We intend to combine probabilistic models with the option for rejection and triggered user interaction in unclear situations.
  4. Co-adaptation of technical systems and the human user: While the former points will mainly be investigated during everyday use of the prosthetic device, the developed methods bear the potential of optimizing the device in the long term towards a true personalized assistant. Currently, the human user still mainly adapts to the prosthesis. We plan to integrate our findings into a co-adaptive process with potentially much better functional outcome.

Outcomes

Demonstration of Transfer Learning

The animated figure shows the Transfer Learning approach applied to EMG data. Each subfigure displays data (small shapes) and model prototypes (large diamonds) for one degree of freedom (DoF). The three classes are negative movement, no movement and positive movement; for example, for the hand opening/closing DoF -1 corresponds to hand closing, 0 corresponds to no movement and +1 corresponds to hand opening. The prototypes have been trained under a stationary distribution. The data displayed stems from a different data distribution after the input data has been disturbed by a shift in the measurement electrodes. The transfer learning algorithm iteratively re-maps the disturbed data to move them closer toward a fitting prototype.

First results have been achieved in the context of Transfer Learning from few data examples. Here, a novel technique based on the maximum likelihood paradigm has been proposed which is optimized to mapping data from a disturbed domain to the original one where a trained classification model is available. This algorithm has been evaluated in the prosthesis control domain where the applied disturbance was a lateral electrode shift. The proposed methodology outperformed several baselines, including a retrained model and a transfer method from the literature. Furthermore, the effects of applying time-dynamical models in the domain of prosthesis control have been investigated systematically and a significant improvement in action recognition has been shown.

Publications