Neural Network Attribution Learning With Sparse Noise Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for extracting attributions from neural network models often include significant noise, making it difficult to interpret the relationships between input and output data, and even noise-reducing methods can complicate attribution interpretation.
Innovation Solution
A learning device and extraction device that calculate attributions based on input and output data, applying restrictions such as sparsity increase to reduce noise and improve interpretation without altering the calculation method, allowing for clearer attribution analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a method that utilizes a value of partial differentiation of an input with respect to an output is used to extract attribution, then it is possible to obtain a relationship between input and output for each sample, but a lot of noise is included in the extracted attribution
Solution Approach 1:
The patent changes the parameter being optimized from attribution extraction accuracy to attribution sparsity. By applying an L1 norm constraint to the attribution during model learning, the patent transforms the optimization objective to produce sparse attributions that are less noisy and more interpretable, while still maintaining the ability to extract input-output relationships for each sample.
Solution Approach 2:
The patent applies different treatment to different parts of the attribution vector by enforcing sparsity constraints. This causes important features to have non-zero attribution values while less important features are suppressed to zero, creating a differentiated quality profile across the attribution elements that reduces noise and improves interpretability.
2Reliability
If a calculation method for an attribution that eliminates noise is used, then noise is reduced, but interpretation of such an attribution per se may be difficult
Solution Approach 1:
The patent changes the optimization parameter from attribution accuracy to attribution sparsity by applying L1 norm constraints. This parameter change produces attributions that are both less noisy and more interpretable, as sparse attributions naturally highlight the most important features while suppressing less relevant ones, making interpretation easier rather than more difficult.
3Reliability
If sparsity is increased as a restriction on attribution, then noise in attribution is reduced, but model complexity increases
Solution Approach 1:
The patent incorporates sparsity constraints (L1 norm) directly into the existing loss function of the neural network model. This approach integrates the complexity management into the standard learning process rather than adding separate complex processing steps, thereby improving attribution quality while minimizing increases in overall system complexity.
Data Source
AI summary
A learning device includes processing circuitry configured to collect a plurality of data, calculate, when inputting the plurality of data as input data to a model and obtaining output data that is output from the model, an attribution that is a degree of contribution of each element of the input data to the output data, based on the input data and the output data, and apply a restriction on the attribution thereto and learn the model.