Cross-Location Activity Classifier Training Without Labeled Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Labeled data for training machine learning models on wearable devices worn at locations other than the wrist is difficult and expensive to obtain, hindering effective activity classification.
Innovation Solution
Utilize a trained activity classifier from a source location with accessible labeled data to generate embeddings that are used to unsupervisedly train a target activity classifier at a different, unlabeled location, leveraging motion sensors at both locations to minimize reconstruction error without labeled data, and optionally using available labeled data for regularization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If labeled data is collected for training activity classifiers at target locations, then training accuracy is improved, but time and expense increase significantly
Solution Approach 1:
The patent applies preliminary action by pre-training an embedding generator at a source location using readily available labeled data. This pre-trained embedding generator is then transferred to the target location, eliminating the need to collect and process labeled data at the target location. The source location's labeled data is utilized in advance to prepare the model components that will be deployed elsewhere, thereby avoiding the time and expense of target location data collection.
Solution Approach 2:
The patent implements copying by transferring the embedding generator from the source location to the target location. Instead of training a new classifier from scratch at the target location with expensive labeled data, the pre-trained embedding generator is copied and adapted. This allows the target location to leverage the source location's training investments while achieving accurate activity classification without requiring target location-specific labeled data.
2Measurement precision
If activity classifiers are trained at multiple locations with labeled data, then location-specific accuracy is improved, but data collection complexity increases
Solution Approach 1:
The patent applies universality by creating a location-agnostic embedding generator that can be deployed at multiple locations. The embedding generator is trained at a source location but designed to be universally applicable to target locations through feature extraction that captures general activity patterns. This single universal model replaces the need for separate location-specific training pipelines, reducing data collection complexity while maintaining location-specific accuracy through the model's adaptive capabilities.
3Ease of manufacture
If unsupervised training is used at target locations, then labeled data requirements are reduced, but training performance may deteriorate
Solution Approach 1:
The patent introduces an intermediary component - the embedding generator - that bridges the gap between source location labeled data and target location unlabeled data. The embedding generator acts as a mediator that translates source location features into target location predictions. This intermediary enables unsupervised training at the target location by leveraging the source location's labeled data indirectly, maintaining training performance without requiring target location labeled data.
Solution Approach 2:
The patent applies parameter changes by transforming the training approach from direct supervised learning to unsupervised learning with embedding transfer. The model parameters are first optimized using labeled data at the source location, then these parameters are transferred and fine-tuned at the target location using unlabeled data. This parameter transformation allows the system to achieve good performance with minimal labeled data at the target location by leveraging the pre-optimized parameters from the source location.
Data Source
AI summary
Disclosed are systems, methods, and apparatus for unsupervised training of an activity classifier for movement measured at a target location of a user using a trained source activity classifier configured to determine an activity based on motion measured at a source location of a user. According to exemplary embodiments, the activity classifier for the target location can be trained using source embeddings generated by a trained classifier for the source location. For example, the source embeddings and the motion measured at the target location may then be provided as training inputs to the target activity classifier for the target location of a user, so that the target activity classifier can reconstruct the source embedding for the target motion without the use of any labeled training data and/or without having to learn the transfer function of the existing trained machine learning model.