Gated Multi-Branch ML Models to Prevent Negative Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training and maintaining multiple machine learning models for various tasks is computationally expensive and often results in compromised inference accuracy due to conflicts between tasks, known as negative transfer or destructive interference.
Innovation Solution
A multi-domain machine learning model is trained using a gating neural network to route inputs to specific neural network blocks, each trained for a particular domain, allowing the model to perform multiple tasks while maintaining desired inference performance by extracting features and using a confusion matrix for clustering and supervision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple machine learning models are trained to perform various tasks, then task-specific inference accuracy is improved, but computational expense and resource utilization increase significantly
Solution Approach 1:
The patent implements a universal machine learning model that can perform multiple tasks through dynamic routing. Instead of maintaining separate models for each task, a single multi-domain model is trained to handle various tasks by routing inputs to appropriate neural network blocks based on task identification, thereby reducing computational resources while maintaining task-specific accuracy
Solution Approach 2:
The patent segments the universal model into multiple domain-specific neural network blocks (e.g., image processing block, text processing block, audio processing block). Each block is specialized for specific tasks, and a gating network dynamically selects which block to activate based on the input task, achieving both efficiency and accuracy
2Use of energy by moving object
If a single machine learning model is trained to perform multiple tasks using transfer learning, then computational expense is reduced, but inference accuracy deteriorates due to negative transfer and destructive interference between tasks
Solution Approach 1:
The patent divides the single model into multiple domain-specific neural network blocks that are activated selectively. This segmentation prevents negative transfer by ensuring that tasks from different domains (e.g., image processing and text processing) do not interfere with each other, as each task is routed to its specialized block
Solution Approach 2:
The patent introduces a gating network as an intermediary that sits between the input and the domain-specific neural network blocks. This gating network identifies the task domain and routes inputs to the appropriate block, acting as a mediator that prevents destructive interference between conflicting tasks while maintaining the efficiency of a single model
Data Source
AI summary
Certain aspects of the present disclosure provide techniques and apparatus for training and inferencing using a multi-domain machine learning model. An example method generally includes extracting, using a first neural network block, a plurality of features associated with inputs in a multi-domain input data set. A confusion matrix is generated based on the extracted plurality of features. A plurality of clusters is identified from the confusion matrix. Each cluster in the plurality of clusters generally corresponds to one or more data domains in the multi-domain input data set. A first gating neural network is trained based on the multi-domain input data set and the identified plurality of clusters. A plurality of second neural network blocks is trained based on a division of the multi-domain input data set into data associated with each cluster of the plurality of clusters.