Inference Model Topology Selection for Latent Bias Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inference models often introduce latent bias due to limited or biased training data, leading to inaccurate and undesirable predictions that negatively impact computer-implemented services.
Innovation Solution
A modified split training method is employed to create a multipath inference model by dividing the existing inference model into a body portion and a head portion, with one path trained to predict the bias feature and the other to maintain predictive power for the target feature, thereby reducing latent bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional inference models are used with limited training data, then model complexity is reduced, but latent bias is introduced leading to inaccurate predictions
Solution Approach 1:
The inference model is divided into multiple independent paths (e.g., main path and auxiliary path), each processing different aspects of the input data. This segmentation allows the model to reduce latent bias by distributing processing responsibilities across multiple specialized pathways rather than relying on a single complex model, thereby improving prediction accuracy without proportionally increasing overall model complexity.
Solution Approach 2:
An intermediary component is introduced between the input data and the final prediction output. This intermediary processes and transforms the input data in a way that reduces latent bias before it reaches the prediction layer, acting as a mediator that improves prediction accuracy by filtering out biased information while maintaining reasonable model complexity.
2Device complexity
If model architecture is simplified to reduce complexity, then training data requirements are reduced, but ability to capture complex patterns decreases
Solution Approach 1:
The model architecture is segmented into multiple specialized paths, each designed to capture specific types of patterns. This segmentation allows the system to maintain low overall complexity while preserving pattern recognition capability, as each simple path focuses on a specific aspect rather than requiring one complex path to handle all patterns.
Solution Approach 2:
The segmented model paths are designed to be universal in their ability to handle different types of input data and patterns. Each path can process various data types within its domain, providing versatile pattern recognition capability without requiring highly specialized complex architectures for each specific pattern type.
3Productivity
If single-path inference models are used, then computational efficiency is improved, but latent bias in predictions increases
Solution Approach 1:
The inference process is segmented into multiple parallel computational paths that can be executed simultaneously. This segmentation enables the system to maintain computational efficiency through parallel processing while reducing latent bias, as each path contributes independently to the final prediction, preventing any single biased path from dominating the output.
Solution Approach 2:
Multiple inference paths are merged at the output stage to produce the final prediction. This merging combines the results from various paths in a way that cancels out individual path biases while preserving computational efficiency, as the merging operation is designed to be computationally lightweight despite integrating multiple path outputs.
Data Source
AI summary
Methods, systems, and devices for providing computer-implemented services are disclosed. To provide the computer-implemented services, inference models used by data processing systems may be managed to reduce the likelihood of the inference models provide inferences indicative of bias features. The inference models may be managed using a divisional process to obtain multipath inference models, as part of a modified split training to reduce mutual information shared with the bias feature. The inferences provided by the inference models may be less likely to include latent bias thereby reducing bias in computer-implemented services provided using the inferences.