Interpretable Machine Learning at Scale Using Locality-Sensitive Hashing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models, particularly deep learning models, are difficult to interpret, hindering their widespread adoption in industries like healthcare and banking due to the lack of transparency and accountability in their decision-making processes.
Innovation Solution
A computer-implemented approach using an interpretable surrogate model that locally approximates the behavior of black-box models by training on a neighborhood of similar inputs, facilitated by locality sensitive hashing to reduce computational costs and enhance interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning models are used to improve performance, then inference accuracy is improved, but model interpretability deteriorates
Solution Approach 1:
The patent introduces an intermediary explanation model that mediates between the complex deep learning model and the user. This explanation model takes the same input as the deep learning model and produces human-interpretable predictions, serving as a translator that preserves the accuracy benefits of deep learning while providing interpretability through a separate, simpler model layer.
Solution Approach 2:
The patent creates a simplified copy or approximation of the deep learning model's behavior. Instead of interpreting the complex internal workings of the deep learning model, the system trains a simpler explanation model that replicates the predictive behavior in an interpretable form, allowing users to understand decision-making without sacrificing accuracy.
2Loss of information
If LIME and SHAP algorithms are used to provide interpretations, then model interpretability is improved, but computational cost increases
Solution Approach 1:
The patent performs preliminary action by pre-training the explanation model using a representative sample of data before actual inference. This allows the explanation model to be ready for rapid deployment without requiring heavy computational resources during the explanation generation phase, as the complex computations are done upfront during training rather than during each individual prediction.
Solution Approach 2:
The patent creates a simplified copy of the deep learning model's predictive behavior that can be executed much more efficiently. The explanation model is trained to replicate the deep learning model's decisions but uses a simpler architecture that requires significantly less computational resources to run, making interpretability accessible at scale.
3Measurement precision
If synthetic data is generated by perturbing features to create local neighborhood, then interpretation accuracy is improved, but data realism deteriorates
Solution Approach 1:
The patent uses a pre-existing dataset with real data points as the foundation for training the explanation model, rather than generating synthetic data by perturbing features. This approach preserves the realism of the data while still providing accurate local neighborhood information for interpretation, as the model learns from actual observed patterns rather than artificial constructions.
Solution Approach 2:
The patent uses a representative sample of real data points that captures the essential patterns of the local neighborhood without requiring exhaustive generation of all possible perturbed variations. This partial approach using real data achieves sufficient interpretation accuracy while maintaining data realism, avoiding the need to generate every possible synthetic variation.
Data Source
AI summary
In systems for interpreting the predictions of a machine learning model with the help of a surrogate model, feature vectors of inputs to the machine learning model can be grouped based on locality sensitive hashes or other hashes that reflect similarity between the feature vectors in matching hash values. For a given prediction to be interpreted and the corresponding input feature vector, a suitable training dataset for the surrogate model can then be obtained at low computational cost by hashing the input feature vector and retrieving stored feature vectors with matching hash values, along with their respective predictions.