Model repair
Concrete Problems in AI Safety — inherited
A prior technique (Ghosh et al. 2016) that alters a trained model after the fact to ensure specified safety properties hold.
Instead of retraining a faulty model, one line of work patches it. Model repair, due to Ghosh and colleagues in 2016, alters a trained model directly until a specified safety property holds, and it enters this corpus as a concrete instance of making machine learning honor a contract.
The sections cover patching versus retraining first, then the technique's role as the ex-post half of a pair.
Patching a trained model instead of retraining it
Model repair, due to Ghosh et al. in 2016, is a technique for altering an already-trained machine learning model directly so that it satisfies a specified safety or correctness property, rather than diagnosing why the property failed and redoing the training process that produced the violation. In practice this means treating the trained model's parameters as adjustable subject to constraints — forcing compliance with a desired property through targeted modification rather than a fresh pass over the training data. The technique appeared the same year as Concrete Problems and is cited as one of two "avenues for obtaining better contracts" on a system's behavior (concrete-problems, §"A unifying view: counterfactual reasoning and machine learning with contracts.", p. 19).
The ex-post half of a pair
Model repair's defining feature, within this corpus, is timing. The edge Machine learning with contracts — is restored post hoc by → Model repair positions it as the response available once a contract violation has already been discovered in a finished model, in contrast to techniques that constrain training or search in advance. That contrast is sharpened by Model repair — is the ex post counterpart to → Reachability analysis, which pairs it directly against Reachability analysis: reachability analysis is ex-ante, constraining the policy-search process itself by requiring a safe region to remain reachable throughout, while model repair is ex-post, altering a model only once training is already complete. Both edges sit in the Before the fact or after: when a safety property gets enforced connective theme, and the second also belongs to Formal verification and control theory recur as safety building blocks. The concept is a child of Machine learning with contracts within the Reframing distributional shift as broken contracts and causal structure theme, and belongs to the Bounding what an agent can change or explore, not just what it's told to want, Where Concrete Problems draws its own lines, and Distributional shift: relax the statistics, or rewrite the contract superthemes — it is, in effect, the corpus's designated answer to the question of what to do once a contract has already broken.