Invertible Factorization Model for Interpretable Signal Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
There is a tradeoff between shallow classifiers, which are easily interpretable but lack accuracy for high-dimensional input signals, and deep classifiers, which achieve high accuracy but are 'black boxes' making it impossible to infer the semantic aspects driving their decisions, limiting their interpretability and robustness.
Innovation Solution
An invertible factorization model is used to determine a latent representation of input signals, allowing for disentangled factors that are continuous and differentiable, enabling the classifier to base decisions on reliable features and providing insight into its decision-making process, thereby improving classification accuracy and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If shallow classifiers are used, then interpretability is improved, but classification accuracy deteriorates for high-dimensional input signals
Solution Approach 1:
The patent segments the deep classifier into an invertible factorization model that decomposes the input signal into disentangled semantic factors. This segmentation allows the classifier to process high-dimensional data through multiple specialized functions while maintaining interpretability by making the decision factors visible and separable.
Solution Approach 2:
The patent introduces latent representations as intermediary variables between the input signal and the final classification. These latent factors serve as mediators that capture semantic aspects of the input while remaining interpretable, bridging the gap between deep learning accuracy and shallow model interpretability.
2Measurement precision
If deep classifiers are used, then classification accuracy is improved, but interpretability deteriorates as the classifier becomes a black box
Solution Approach 1:
The patent applies inversion by creating an invertible factorization model where the transformation from input to latent representation can be reversed. This allows the system to not only forward-process data for accurate classification but also to trace back and interpret which semantic factors contributed to the decision, effectively opening the black box.
Solution Approach 2:
The patent changes the parameter representation by transforming the raw high-dimensional input into a structured latent space with disentangled factors. This parameter transformation maintains the accuracy benefits of deep learning while organizing the information in an interpretable manner that reveals semantic meaning.
3Measurement precision
If deep classifiers are used, then classification accuracy is improved, but robustness to perturbations deteriorates due to reliance on unreliable features
Solution Approach 1:
The patent extracts and isolates specific semantic factors from the input signal through the invertible factorization model. By taking out and separately representing individual semantic aspects in the latent space, the classifier can focus on reliable features while ignoring noise and irrelevant variations, thereby improving robustness.
Solution Approach 2:
The patent applies local quality by ensuring that each latent factor represents a specific semantic aspect with high fidelity. This localized representation means that each factor captures a particular property of the input signal, making the classification more robust to perturbations in other areas while maintaining sensitivity to relevant changes.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Computer-implemented method for determining an output signal (y) for an input signal (x) using a classifier (60), wherein the output signal (y) characterizes a classification of the input signal (x), the method comprising the steps of: • Determining a latent representation (z) based on the input signal (x) by means of an invertible factorization model (61) comprised in the classifier, wherein the latent representation (z) comprises a plurality of factors (z1, z2, z3, z4) and wherein the invertible factorization model (61) is characterized by: - A plurality of functions (F1, F2, F3, F4), wherein any one function (F1, F2, F3, F4) from the plurality of functions is continuous and almost everywhere continuously differentiable, wherein the function (F1, F2, F3, F4) is further configured to accept either the input signal (x) or at least one factor (z1, z2, z3, z4) provided by another function (F1, F2, F3, F4) of the plurality of functions (F1, F2, F3, F4) as an input and wherein the function (F1, F2, F3, F4) is further configured to provide at least one factor (z1, z2, z3, Z4), wherein the at least one factor (z1, z2, z3, z4) is either provided as at least part of the latent representation (z) or is provided as at least part of an input of another function (F1, F2, F3, F4) of the plurality of functions, wherein there exists an inverse function that corresponds with the function (F1, F2, F3, F4), is continuous, is almost everywhere continuously differentiable and is configured to determine the input of the layer based on the at least one factor (z1, z2, z3, z4) provided from the function (F1, F2, F3, F4); • Determining the output signal (y) based on the latent representation (z) by means of an internal classifier (62) comprised in the classifier (60).