Neural Network Input Feature Distillation via Mask Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for attributing input features to the outputs of deep neural networks face challenges due to the non-linear structure of modern DNNs and the lack of ground truth for quantitatively measuring attribution reliability.
Innovation Solution
A method and device for distilling input features through an artificial neural network model, which involves comparing outputs with inputs to extract local attributions, using a mask extractor to filter noise and irrelevant features, and aggregating attributions to produce reliable feature contributions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If input attribution methods are used to measure feature contribution, then the relative importance of input features can be identified, but the reliability of attribution remains low due to non-linear DNN structure and lack of ground truth
Solution Approach 1:
The patent segments the input features into multiple groups and performs attribution analysis on each segment separately. By dividing the input into segments and calculating attributions for each segment, the method reduces the complexity of tracking relationships in the non-linear DNN structure and improves measurement precision of feature contributions.
Solution Approach 2:
The patent introduces an intermediary attribution distribution that serves as a mediator between the input features and the model output. This intermediary representation allows for more reliable attribution by providing a structured way to track feature contributions through the non-linear transformations, addressing the lack of ground truth problem.
2Reliability
If the DNN model becomes larger and more complex, then the model's predictive capability improves, but the ability to investigate and understand internal decision-making mechanisms deteriorates
Solution Approach 1:
The patent extracts attribution information from the complex DNN model by calculating the sensitivity of the output to each input feature. This extraction process isolates the decision-making mechanisms from the overall model complexity, allowing investigation and analysis of individual feature contributions even in large-scale models.
Solution Approach 2:
By segmenting the model's decision-making process into individual feature attributions, the patent makes the internal mechanisms of complex DNNs investigable. Each feature's contribution is independently analyzed, breaking down the black-box complexity into interpretable components.
3Measurement precision
If attribution methods are applied to non-linear DNN structures, then feature importance can be measured, but the difficulty of correctly tracking input-output relationships increases
Solution Approach 1:
The patent uses an intermediary attribution distribution as a mediator to track input-output relationships in non-linear DNNs. This intermediary structure simplifies the tracking process by providing a systematic way to propagate and measure feature contributions through the non-linear transformations, reducing the difficulty of measurement.
Solution Approach 2:
The patent employs feedback mechanisms where the attribution distribution is updated based on the model's output sensitivity. This feedback loop allows for iterative refinement of the attribution measurement, improving precision while managing the complexity of tracking relationships in non-linear structures.
Data Source
AI summary
Provided are a device and method for detecting input features of an artificial neural network model. The method includes comparing an output of an artificial neural network model with a first input to extract a first local attribution for a feature of the first input, extracting a portion of the first local attribution of which an absolute value is a threshold or more using a mask extractor, and updating the first input with a second input by applying an extraction result to the first input.