ML Models with Integrated Uncertainty for Activity Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models for activity recognition, particularly in open-world environments, face challenges in distinguishing between activities due to sensor noise and variations, leading to incorrect predictions and insufficient discrimination between classes.
Innovation Solution
Implementing machine learning models with integrated uncertainty using a complete Bayesian formulation for feature extraction and prediction, along with metric-based learning and temporal smoothing, and model compression techniques such as Shapley value-based feature selection to improve prediction accuracy and resource efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional machine learning models are used for activity recognition, then the models can make predictions from input data, but the models fail to provide reliability information and produce incorrect predictions under sensor noise and variations
Solution Approach 1:
The patent transforms deterministic model parameters into probabilistic distributions, changing the parameter representation from fixed values to uncertainty-aware distributions. This allows the model to output both predictions and reliability measures, resolving the contradiction between prediction capability and reliability information provision.
Solution Approach 2:
The patent introduces uncertainty quantification as an intermediary layer between the traditional ML model and the prediction output. This intermediary provides reliability information without compromising the original prediction function, enabling the system to distinguish between confident and uncertain predictions.
2Reliability
If feature extraction and prediction are performed separately in traditional ML models, then the processing is straightforward, but the models lack robustness against unknown classes and sensor variations
Solution Approach 1:
The patent merges feature extraction and prediction into a unified Bayesian framework where both operations are performed simultaneously with integrated uncertainty propagation. This integration improves robustness by treating features and predictions as jointly probabilistic, though it increases formulation complexity.
Solution Approach 2:
The Bayesian formulation serves multiple functions simultaneously: it performs feature extraction, prediction, and uncertainty quantification in a single unified model. This multi-functionality provides robustness against unknown classes while managing complexity through a comprehensive probabilistic framework.
3Measurement precision
If comprehensive feature sets are used to improve activity recognition accuracy, then the discrimination between highly correlated activities improves, but the computational power and resource requirements increase
Solution Approach 1:
The patent extracts and focuses only on the most discriminative features through Bayesian feature selection, removing redundant features that contribute to computational overhead. This extraction maintains discrimination precision for highly correlated activities while reducing the computational burden of processing comprehensive feature sets.
Solution Approach 2:
The patent changes the parameter representation to probabilistic distributions, which enables more efficient computation through analytical solutions and approximations. This parameter transformation maintains discrimination capability while reducing computational energy consumption through optimized probabilistic inference.
4Loss of information
If traditional ML models are trained with labeled data, then the models can learn from supervised signals, but the models lack confidence scores and cannot indicate prediction reliability
Solution Approach 1:
The Bayesian model provides feedback in the form of confidence scores and uncertainty measures that indicate prediction reliability. This feedback mechanism allows the system to self-assess the quality of its predictions without requiring additional labeled data, thus recovering lost confidence information efficiently.
Solution Approach 2:
The probabilistic framework enables the model to self-assess its own prediction reliability through intrinsic uncertainty quantification. The model serves itself by generating confidence scores as a natural byproduct of the Bayesian inference process, without requiring external verification or additional data processing.
Data Source
AI summary
A system includes memory and a processing device, operatively coupled to the memory, to obtain an input signal corresponding to data obtained from a data source, extract a set of features using the input signal, generate a set of feature tracking data from the set of features, compress a machine learning model to obtain a compressed model by identifying a subset of features based on the set of tracking data, and use the compressed model to make a prediction based on the set of feature tracking data. The set of features includes a set of confidence features and a set of uncertainty features, and the set of feature tracking data includes a set of confidence feature tracking data and a set of uncertainty feature tracking data.