Search for: All records

Machine unlearning---efficiently removing the effect of a small "forget set" of training data on a pre-trained machine learning model---has recently attracted significant research interest. Despite this interest, however, recent work shows that existing machine unlearning techniques do not hold up to thorough evaluation in non-convex settings. In this work, we introduce a new machine unlearning technique that exhibits strong empirical performance even in such challenging settings. Our starting point is the perspective that the goal of unlearning is to produce a model whose outputs are statistically indistinguishable from those of a model re-trained on all but the forget set. This perspective naturally suggests a reduction from the unlearning problem to that of *data attribution, where the goal is to predict the effect of changing the training set on a model's outputs. Thus motivated, we propose the following meta-algorithm, which we call Datamodel Matching (DMM): given a trained model, we (a) use data attribution to predict the output of the model if it were re-trained on all but the forget set points; then (b) fine-tune the pre-trained model to match these predicted outputs. In a simple convex setting, we show how this approach provably outperforms a variety of iterative unlearning algorithms. Empirically, we use a combination of existing evaluations and a new metric based on the KL-divergence to show that even in non-convex settings, DMM achieves strong unlearning performance relative to existing algorithms. An added benefit of DMM is that it is a meta-algorithm, in the sense that future advances in data attribution translate directly into better unlearning algorithms, pointing to a clear direction for future progress in unlearning. 

We study the data deletion problem for convex models. By leveraging techniques from convex optimization and reservoir sampling, we give the first data deletion algorithms that are able to handle an arbitrarily long sequence of adversarial updates while promising both per-deletion run-time and steady-state error that do not grow with the length of the update sequence. We also introduce several new conceptual distinctions: for example, we can ask that after a deletion, the entire state maintained by the optimization algorithm is statistically indistinguishable from the state that would have resulted had we retrained, or we can ask for the weaker condition that only the observable output is statistically indistinguishable from the observable output that would have resulted from retraining. We are able to give more efficient deletion algorithms under this weaker deletion criterion. 

We consider settings in which the right notion of fairness is not captured by simple mathematical definitions (such as equality of error rates across groups), but might be more complex and nuanced and thus require elicitation from individual or collective stakeholders. We introduce a framework in which pairs of individuals can be identified as requiring (approximately) equal treatment under a learned model, or requiring ordered treatment such as "applicant Alice should be at least as likely to receive a loan as applicant Bob". We provide a provably convergent and oracle efficient algorithm for learning the most accurate model subject to the elicited fairness constraints, and prove generalization bounds for both accuracy and fairness. This algorithm can also combine the elicited constraints with traditional statistical fairness notions, thus "correcting" or modifying the latter by the former. We report preliminary findings of a behavioral study of our framework using human-subject fairness constraints elicited on the COMPAS criminal recidivism dataset. 

Kearns, Neel, Roth, and Wu [ICML 2018] recently proposed a notion of rich subgroup fairness intended to bridge the gap between statistical and individual notions of fairness. Rich subgroup fairness picks a statistical fairness constraint (say, equalizing false positive rates across protected groups), but then asks that this constraint hold over an exponentially or infinitely large collection of subgroups defined by a class of functions with bounded VC dimension. They give an algorithm guaranteed to learn subject to this constraint, under the condition that it has access to oracles for perfectly learning absent a fairness constraint. In this paper, we undertake an extensive empirical evaluation of the algorithm of Kearns et al. On four real datasets for which fairness is a concern, we investigate the basic convergence of the algorithm when instantiated with fast heuristics in place of learning oracles, measure the tradeoffs between fairness and accuracy, and compare this approach with the recent algorithm of Agarwal, Beygelzeimer, Dudik, Langford, and Wallach [ICML 2018], which implements weaker and more traditional marginal fairness constraints defined by individual protected attributes. We find that in general, the Kearns et al. algorithm converges quickly, large gains in fairness can be obtained with mild costs to accuracy, and that optimizing accuracy subject only to marginal fairness leads to classifiers with substantial subgroup unfairness. We also provide a number of analyses and visualizations of the dynamics and behavior of the Kearns et al. algorithm. Overall we find this algorithm to be effective on real data, and rich subgroup fairness to be a viable notion in practice 

Data that is gathered adaptively --- via bandit algorithms, for example --- exhibits bias. This is true both when gathering simple numeric valued data --- the empirical means kept track of by stochastic bandit algorithms are biased downwards --- and when gathering more complicated data --- running hypothesis tests on complex data gathered via contextual bandit algorithms leads to false discovery. In this paper, we show that this problem is mitigated if the data collection procedure is differentially private. This lets us both bound the bias of simple numeric valued quantities (like the empirical means of stochastic bandit algorithms), and correct the p-values of hypothesis tests run on the adaptively gathered data. Moreover, there exist differentially private bandit algorithms with near optimal regret bounds: we apply existing theorems in the simple stochastic case, and give a new analysis for linear contextual bandits. We complement our theoretical results with experiments validating our theory. 

Settings such as lending and policing can be modeled by a centralized agent allocating a scarce resource (e.g. loans or police officers) amongst several groups, in order to maximize some objective (e.g. loans given that are repaid, or criminals that are apprehended). Often in such problems fairness is also a concern. One natural notion of fairness, based on general principles of equality of opportunity, asks that conditional on an individual being a candidate for the resource in question, the probability of actually receiving it is approximately independent of the individual’s group. For example, in lending this would mean that equally creditworthy individuals in different racial groups have roughly equal chances of receiving a loan. In policing it would mean that two individuals committing the same crime in different districts would have roughly equal chances of being arrested. In this paper, we formalize this general notion of fairness for allocation problems and investigate its algorithmic consequences. Our main technical results include an efficient learning algorithm that converges to an optimal fair allocation even when the allocator does not know the frequency of candidates (i.e. creditworthy individuals or criminals) in each group. This algorithm operates in a censored feedback model in which only the number of candidates who received the resource in a given allocation can be observed, rather than the true number of candidates in each group. This models the fact that we do not learn the creditworthiness of individuals we do not give loans to and do not learn about crimes committed if the police presence in a district is low.