Exam-style
A lender's regulator requires that every declined application reach the applicant with the reasons, in terms the applicant can act on. Four candidates: a logistic regression, a gradient-boosted tree with per-feature attribution, a deep neural network, and a foundation model on Amazon Bedrock. Which choice meets the requirement?
Reveal the answer
C. The logistic regression, or the gradient-boosted tree with its per-feature attribution computed and stored at decision time where the accuracy gain justifies the extra machinery
Interpretability is understanding a model’s internal mechanics. Explainability is a post-hoc account of one output, produced without opening the model. A logistic regression gives both: each feature has a coefficient, and one applicant’s contribution is that coefficient times their feature value. A gradient-boosted tree gives the second, through per-decision tree SHAP attribution. That machinery is bolted on afterwards, and it survives an audit three years later only if the attribution values, the feature values and the model version were written down at decision time. A deep neural network gives no interpretability. Integrated gradients would give it per-prediction attribution, but neither deep-network option here produces that. Global feature importance describes the portfolio, not this applicant. A person reconstructing a story about a score describes their own reasoning. Asking the foundation model for its own reasons is the tempting move and the worst one. There are no stable input features to attribute, so the paragraph is generated alongside the decision rather than derived from it. At best a retrieval-backed model cites the sources it retrieved. The last option is where tradeoffs between model safety and transparency bite: publishing the coefficients and the cut-off hands anyone seeking approval a recipe for clearing the threshold. Disclose enough to be accountable, and no more.
Q. Every declined application has to name its reasons. Logistic regression, boosted tree with attribution, deep network, or a Bedrock foundation model?
A. The logistic regression, or the boosted tree with per-feature attribution stored at decision time. Both sides of that trade get a number.
Why? AWS asks you to describe the differences between models that are transparent and explainable and models that are not. Interpretability is reading the internal mechanics: coefficients, split points, the score as arithmetic. Explainability accounts for one output after the fact, without opening the model. The regression gives both; the boosted tree gives the second, through attribution stored per decision. Neither deep-network option produces per-decision attribution, though integrated gradients could: a global importance chart describes the portfolio, and a human’s account describes the human. A foundation model asked to justify its own decline outputs a plausible paragraph, not a report of the arithmetic. What it can genuinely deliver is traceability to the passages it retrieved. The guide’s phrase for weighing lost accuracy against gained transparency is measure interpretability and performance. Write down the accuracy each option gives up and the explanation it can produce. Publishing the whole model satisfies transparency and tells an applicant which number to move, so those tradeoffs resolve at disclosing enough to be accountable.