Anyone in biomechanics research involved in human activity recognition?
One interesting AI approach is "Few-shot Learning" or "Zero-shot Learning" where a model can identify classes that are 'new' or 'unseen' from the training set. Could this approach be used in a biomechanics context?
Imagine you are wearing an exoskeleton and have to move your foot in an awkward position to avoid falling. However, the machine learning-based controller is confused about what you're trying to do because the training data of the model does not include this scenario's sequence of actions.
Humans often perform tasks that perhaps were 'unexpected.' It's often difficult to capture every type of activity a human may encounter in daily life in a motion-capture setting. Could Zero-shot Learning be the key ingredient to predict an activity that the model never previously encountered?
Here's what I am thinking:
David Stephen Hollinger It's intriguing that you're considering using zero-shot learning for human activity identification in the context of biomechanics research. Zero-shot learning is a sort of machine learning in which a model can categorize new classes of data that it has never seen before using knowledge from previous classes. This differs from typical machine learning, in which a model can only categorize new data if it has already encountered comparable samples in its training set.
You are accurate that few-shot learning techniques like Prototypical Networks have been employed in image identification tasks, and they may be useful in biomechanics datasets. The model would learn a feature space that captures the underlying structure of the data in this situation.
It is crucial to highlight, however, that developing a biomechanics dataset adequate for zero-shot learning would be difficult. Human activity identification is a hard problem, and it would be impossible to capture all potential actions that a human would face in daily life. There are, however, initiatives underway to create big datasets for activity recognition, such as the Human Activity Recognition Using Smartphones Data Set (HAR) and Kinetics, which capture a wide range of human activities.
Furthermore, the size of the dataset may influence the Zero-shot learning strategy. A larger dataset with more samples and classes might help the model generalize better. Also, adequate annotation is required because the model will utilize it as a 'word embedding' to categorize.
It is plausible that zero-shot learning may be used to anticipate unobserved events in a biomechanics environment, but this would need extensive data gathering and annotation. It would also necessitate the development of fresh technologies capable of dealing with the complexities of human activity recognition.
It would be useful to run some proof-of-concept experiments to assess the possibility of employing zero-shot learning for human activity recognition in a biomechanics setting, as well as to compare its performance to other approaches.
Dear David Stephen Hollinger ,
Human activity recognition aims to recognize human activities from sensor readings. Most of existing methods in this area can only recognize activities contained in training dataset.
Conference Paper Zero-shot human activity recognition via nonlinear compatibi...
Regards,
Shafagat
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification