Hessian-Free Sample-Based Model Explanations with TracIn++
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing sample-based model explanation methods for natural language processing, such as Influence Functions, face challenges with high computing costs due to Hessian matrix approximation and may not be faithful to the model, leading to spurious explanations, especially for large-scale datasets.
Innovation Solution
A Hessian matrix-free approach using TracIn++ that calculates importance scores through gradient perturbations of training examples, ensuring faithfulness by maintaining a set of faithful model variants close to the final model, and employing span-based influence functions for improved interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If Influence Functions are used for sample-based model explanations, then explanation faithfulness is improved, but computing cost increases prohibitively due to Hessian matrix approximation
Solution Approach 1:
The patent extracts and removes the computationally expensive Hessian matrix approximation step from the Influence Functions methodology. By using alternative mathematical formulations that compute influence scores through gradient-based methods rather than Hessian-based methods, the patent eliminates the primary source of computational overhead while preserving the core functionality of explaining model decisions through training example influence.
Solution Approach 2:
The patent creates simplified copies of the Influence Functions approach by developing alternative influence score calculations that mimic the faithfulness properties of Hessian-based methods but use computationally lighter operations. These copied methods replicate the essential explanatory power without requiring the expensive matrix inversions and approximations of the original approach.
2Loss of energy
If TracIn is used for Hessian-free influence scoring, then computing cost is reduced, but explanation faithfulness deteriorates due to spurious explanations from sub-optimal checkpoints
Solution Approach 1:
The patent modifies the TracIn methodology by changing key parameters and formulations in the influence score calculation. Specifically, it adjusts the mathematical expressions used to compute training example influence to better align with the final model's decision boundaries, thereby improving faithfulness while maintaining the computational efficiency that makes TracIn attractive compared to traditional Influence Functions.
3Measurement precision
If large-scale datasets are used with modern NLP models, then model performance is improved, but explanation efficiency deteriorates due to the extensive size of training data
Solution Approach 1:
The patent segments the large-scale training dataset into manageable batches or subsets for influence score computation. By processing training examples in smaller groups rather than evaluating all training data simultaneously, the patent reduces the computational burden of generating explanations while still capturing the collective influence of the entire training set on model decisions.
Data Source
AI summary
Hessian matrix-free sample-based techniques for model explanations that are faithful to the model are provided. In one aspect, a method for explaining a machine learning model {circumflex over (θ)} (e.g., for natural language processing) is provided. The method includes: training the machine learning model {circumflex over (θ)} with training data D; obtaining a decision of the machine learning model {circumflex over (θ)}; and explaining the decision of the machine learning model {circumflex over (θ)} using training examples from the training data D.