Faithful Perturbations and Evaluations for Post-Hoc Local Explanation Methods
Bibliographic record
Abstract
There are many pre-trained AI models currently in use because of their high performance and quick responses on important tasks. With new laws and legislature passing that is aware and critical of these methods we now require explanations from these models. However, most of the original training data are no longer available and it's impractical to expend resources to train a new explainable model. An alternative is to produce explanations of these pre-trained models, known as post-hoc explanations. The most popular of these methods, LIME, has seen a great deal of employment on this problem, but it has unaddressed issues with performance. Something we presume is caused by the quality of the local data generated to train it, known as perturbations. To fix this issue and clarify the goals of post-hoc explanations we propose using a distribution better follows patterns in the data when generating perturbed samples. A distribution is used to add noise or create more local samples, producing perturbations. Originally this is the normal distribution but there are many others. We evaluate each approach by estimating fidelity on real local-data using the nearest neighbors to the explained sample. We find that when there are correlations between features, using multivariate perturbations greatly improves generalizability, more so than other perturbation approaches. This works particularly well for complex post-hoc explainers, and when there is little to gain from interactions there is no noticeable decrease to performance to worry about.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.001 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".