Explainability (ML Forensic)
Definition
The capacity of a machine-learning model to articulate, in interpretable terms, the features driving its classification or estimation output. Required under Daubert, UK Criminal Procedure Rules Part 19, and the analogous standards in India and Australia, where an expert must explain the basis for their opinion. Neural networks require explainability tools (LIME, SHAP) to partially achieve this.
- Legal drivers
- Daubert (US), UK CrimPR Part 19, India/Australia analogues
- Explainability tools
- LIME, SHAP, for post-hoc interpretation of neural nets
- Applies to
- ML classification or estimation used in expert opinion
- Limitation
- LIME/SHAP give partial, approximate explanations only
Common questions
Why do LIME and SHAP only partially achieve explainability for neural networks?+
They approximate a complex model's local behaviour around a specific prediction rather than exposing its actual internal decision process, showing which input features likely influenced one output, not a complete, verifiable account of how the network reached its conclusion.
Would a court reject an ML-based opinion outright for using a neural network?+
Not automatically. Admissibility standards like Daubert ask whether the method's basis can be tested, its error rate known, and its reasoning explained to the fact-finder, so a validated network paired with explanation tools can be admissible.
Related terms
- Agisoft Metashape
- Commercial photogrammetry software (Agisoft LLC, Saint Petersburg) widely used in forensic anthropology for 3D model production from image sets. Accuracy of approximately...
- COLMAP
- Open-source SfM photogrammetry software maintained by ETH Zurich, providing zero-cost 3D reconstruction from image sets. Used in under-resourced forensic laboratories in South...
- DBSML
- Detroit-Bone Sex/Maturity Logic: a logistic-regression-based classification system for sex estimation from cranial and postcranial linear measurements, developed at Wayne State University, providing...
- Geomorph (R Package)
- An open-source R package (Adams and Otarola-Castillo 2013; actively maintained through 2024) implementing Procrustes superimposition, principal component analysis in shape space, multivariate...
- LASSO Regression
- Least Absolute Shrinkage and Selection Operator: a regularised regression method that shrinks non-predictive coefficients toward zero, performing simultaneous feature selection and regression....
- Micro-CT
- Computed tomography at voxel sizes of 10 to 100 micrometres, resolving individual trabeculae and bone microarchitecture invisible to clinical CT or surface...
- Procrustes Superimposition
- The statistical procedure in geometric morphometrics that removes non-shape variation (translation, rotation, scale) from a set of landmark coordinates, leaving only shape...
- Structure-from-Motion (SfM)
- A photogrammetric technique in which overlapping images from a drone or camera, processed by software, generate a three-dimensional point cloud and textured...
- Thin-Plate Spline (TPS)
- A mathematical warping function used in geometric morphometrics to visualise the shape difference between a reference specimen and a target, deforming a...