Hybrid ML Architecture Combining PGM and Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning approaches, such as Neural Networks, struggle with classifying data that is context-dependent and sparse, as seen in naval tracking data, where common models fail to efficiently capture intricate behaviors and require relearning for each context, making it challenging to interpret and classify.
Innovation Solution
A method is developed that combines Probabilistic Graphical Models with machine learning models through embedding training, where extension variables are added to the model to capture context-specific dependencies, enabling efficient learning and classification across different contexts without relearning the entire model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If Probabilistic Graphical Models are used to model context and causal relations, then the ability to efficiently model context and incorporate expert knowledge is improved, but the size and complexity of the model grows with the number of relations and states of variables
Solution Approach 1:
The patent segments the complex PGM into multiple smaller PGMs, each representing a specific context or aspect of the data. Each segmented model has reduced complexity while collectively they cover the full range of relations and states. This allows the system to maintain high adaptability to different contexts without requiring any single model to be overly complex.
Solution Approach 2:
The patent implements dynamic model selection where the system automatically selects or switches between different PGMs based on the input data characteristics and context. This dynamic approach allows the system to adapt to different contexts using appropriate pre-trained models rather than using a single large static model, thereby reducing the effective complexity for any given task.
2Measurement precision
If Probabilistic Graphical Models are extended to capture intricate behaviors with higher modelling resolution, then the ability to capture intricate behaviors is improved, but the learning of such models becomes challenging
Solution Approach 1:
The patent applies preliminary action by pre-training multiple PGMs on different datasets and contexts before deployment. Each PGM is pre-trained to capture specific intricate behaviors and patterns. During inference, the system selects the pre-trained model that best matches the current data characteristics, avoiding the need to learn high-resolution models from scratch for each new context and thereby reducing learning difficulty.
3Productivity
If common machine learning approaches like Neural Networks are used for classification, then the ability to process rich patterns is improved, but the ability to efficiently handle context-dependent sparse data deteriorates
Solution Approach 1:
The patent creates a universal classification system using multiple PGMs that can handle both rich pattern data and context-dependent sparse data. Each PGM is designed to be multi-functional, capable of processing different data types and contexts. The system selects the appropriate PGM based on data characteristics, providing universal functionality that adapts to both image-like rich patterns and sparse contextual data equally effectively.
Data Source
AI summary
A method of building a computer implemented data classifier for classifying data from a certain context is provided, whereby the classifier is based on a model obtained by transfer learning combining Probabilistic Graphical Models (PGM) and arbitrary, context independent machine learned models enabled by special modelling patterns, where variables representing outputs of machine learned models are added to the PGM.