Guided Latent Feature Space for Controlled ML Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional AI training technologies lack the ability for data scientists to control and specify the latent features learned during training, which adversely impacts model performance and practical application.
Innovation Solution
Implement a system that specifies and guides the contributions of latent feature space through expert knowledge, using auxiliary tasks to generate and transfer latent features to a target model, with constrained contribution levels based on design requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If conventional AI training technologies are used, then the model can be trained automatically based on training data, but data scientists have limited ability to control or specify the latent features learned during training
Solution Approach 1:
The patent introduces auxiliary tasks as intermediary components that mediate between the training data and the target model. These auxiliary tasks generate auxiliary latent features that serve as controlled intermediaries, allowing data scientists to specify and control certain latent features while maintaining automatic training. The auxiliary tasks act as a bridge that translates design requirements into controlled latent feature contributions without eliminating automation.
Solution Approach 2:
The patent segments the latent feature space into two distinct components: auxiliary latent features (from auxiliary tasks) and residual latent features (from the target model). This segmentation allows independent control and specification of different feature subsets. By dividing the feature space, the system enables selective control over specific latent features while maintaining automatic training for the overall model.
2Device complexity
If the latent feature space is fully data-driven without control, then the training process is simple and automatic, but the model performance and practical application are adversely impacted
Solution Approach 1:
The patent applies local quality by allowing different parts of the latent feature space to have different degrees of control. Specifically, auxiliary latent features from selected auxiliary tasks are subjected to controlled contribution levels (e.g., L2-norm constraints), while other residual latent features remain freely data-driven. This localized control approach improves model reliability for specific critical features without unnecessarily complicating the entire training process.
Solution Approach 2:
The patent changes the parameter control mechanism by introducing contribution level parameters (such as L2-norm constraints) that regulate the influence of auxiliary latent features. By adjusting these parameters, the system can flexibly control the degree of influence auxiliary features have on the target model, thereby improving performance without requiring complete restructuring of the training process.
3Ease of operation
If multiple auxiliary tasks are trained to generate controlled latent features, then control over latent feature contributions is achieved, but the training time and computational resources increase
Solution Approach 1:
The patent applies partial action by selecting only a specific subset of auxiliary tasks rather than training all possible auxiliary tasks. The system identifies and trains only those auxiliary tasks that are relevant to the design requirements and contribute most value to the target model. This selective approach achieves necessary control over latent features while minimizing the time and computational resources required compared to training all auxiliary tasks.
4Reliability
If contribution levels of auxiliary latent features are constrained, then alignment with business objectives is improved, but the model flexibility and adaptability are reduced
Solution Approach 1:
The patent segments the latent features into controlled auxiliary features and flexible residual features. The auxiliary latent features from selected auxiliary tasks are constrained to align with business objectives, while the residual latent features remain unconstrained and adaptable. This segmentation preserves model flexibility by allowing the residual features to adapt to new patterns and data while maintaining reliability through controlled auxiliary features.
Solution Approach 2:
The patent introduces dynamic control through contribution level parameters that can be adjusted based on specific needs. The L2-norm constraints on auxiliary latent features are not fixed but can be modified to balance alignment with business objectives and model adaptability. This dynamic parameter adjustment allows the system to flexibly respond to different requirements while maintaining controlled influence of auxiliary features.
Data Source
AI summary
Systems, methods and products for quantitative translation of design requirements into a machine learning framework for training a classification model. A plurality of auxiliary tasks associated with a plurality of auxiliary task models are specified. The plurality of auxiliary task models are concurrently trained on the auxiliary tasks to generate one or more latent features learned by the plurality of auxiliary task models. The one or more latent features may be transferred from the plurality of auxiliary task models to augment a latent feature space of a target task for the classification model. Contribution levels of the transferred one or more latent features are adjusted based on design requirements for the target task for the classification model. First and second contribution levels are specified for respective first and second sets of auxiliary task latent features being quantified and enforced.