A few large, homogenous, pre-trained models undergird many machine learning systems -and often, these models contain harmful stereotypes learned from the internet.We investigate the bias transfer hypothesis: the theory that social biases (such as stereotypes) internalized by large language models during pre-training transfer into harmful task-specific behavior after fine-tuning.For two classification tasks, we find that reducing intrinsic bias with controlled interventions before finetuning does little to mitigate the classifier's discriminatory behavior after fine-tuning.Regression analysis suggests that downstream disparities are better explained by biases in the fine-tuning dataset.Still, pre-training plays a role: simple alterations to co-occurrence rates in the fine-tuning dataset are ineffective when the model has been pre-trained.Our results encourage practitioners to focus more on dataset quality and context-specific harms.

We consider the problem of utility maximization in online ranking applications while also satisfying a pre-defined fairness constraint. We consider batches of items which arrive over time, already ranked using an existing ranking model. We propose online post-processing for re-ranking these batches to enforce adherence to the pre-defined fairness constraint, while maximizing a specific notion of utility. To achieve this goal, we propose two deterministic re-ranking policies. In addition, we learn a re-ranking policy based on a novel variation of learning to search. Extensive experiments on real world and synthetic datasets demonstrate the effectiveness of our proposed policies both in terms of adherence to the fairness constraint and utility maximization. Furthermore, our analysis shows that the performance of the proposed policies depends on the original data distribution w.r.t the fairness constraint and the notion of utility.

The prevailing wisdom is that a model's fairness and its accuracy are in tension with one another. However, there is a pernicious {\em modeling-evaluating dualism} bedeviling fair machine learning in which phenomena such as label bias are appropriately acknowledged as a source of unfairness when designing fair models, only to be tacitly abandoned when evaluating them. We investigate fairness and accuracy, but this time under a variety of controlled conditions in which we vary the amount and type of bias. We find, under reasonable assumptions, that the tension between fairness and accuracy is illusive, and vanishes as soon as we account for these phenomena during evaluation. Moreover, our results are consistent with an opposing conclusion: fairness and accuracy are sometimes in accord. This raises the question, {\em might there be a way to harness fairness to improve accuracy after all?} Since most notions of fairness are with respect to the model's predictions and not the ground truth labels, this provides an opportunity to see if we can improve accuracy by harnessing appropriate notions of fairness over large quantities of {\em unlabeled} data with techniques like posterior regularization and generalized expectation. Indeed, we find that semi-supervision not only improves fairness, but also accuracy and has advantages over existing in-processing methods that succumb to selection bias on the training set.

We propose a novel Stackelberg game model of MDP interdiction in which the defender modifies the initial state of the planner, who then responds by computing an optimal policy starting with that state. We first develop a novel approach for MDP interdiction in factored state space that allows the defender to modify the initial state. The resulting approach can be computationally expensive for large factored MDPs. To address this, we develop several interdiction algorithms that leverage variations of reinforcement learning using both linear and non-linear function approximation. Finally, we extend the interdiction framework to consider a Bayesian interdiction problem in which the interdictor is uncertain about some of the planner's initial state features. Extensive experiments demonstrate the effectiveness of our approaches.

Computational protein design has been successful in modeling fixed backbone proteins in a single conformation. However, when modeling large ensembles of flexible proteins, current methods in protein design have been insufficient. Large barriers in the energy landscape are difficult to traverse while redesigning a protein sequence, and as a result current design methods only sample a fraction of available sequence space. We propose a new computational approach that combines traditional structure-based modeling using the Rosetta software suite with machine learning and integer linear programming to overcome limitations in the Rosetta sampling methods. We demonstrate the effectiveness of this method, which we call BROAD, by benchmarking the performance on increasing predicted breadth of anti-HIV antibodies. We use this novel method to increase predicted breadth of naturally-occurring antibody VRC23 against a panel of 180 divergent HIV viral strains and achieve 100% predicted binding against the panel. In addition, we compare the performance of this method to state-of-the-art multistate design in Rosetta and show that we can outperform the existing method significantly. We further demonstrate that sequences recovered by this method recover known binding motifs of broadly neutralizing anti-HIV antibodies. Finally, our approach is general and can be extended easily to other protein systems. Although our modeled antibodies were not tested in vitro, we predict that these variants would have greatly increased breadth compared to the wild-type antibody.

Drug and vaccination therapies are important tools in the battle against infectious diseases such as HIV and influenza. However, many viruses, including HIV, can rapidly escape the therapeautic effect through a sequence of mutations. We propose to design vaccines, or, equivalently, antibody sequences that make such evasion difficult. We frame this as a bilevel combinatorial optimization problem of maximizing the escape cost, defined as the minimum number of virus mutations to evade binding an antibody. Binding strength can be evaluated by a protein modeling software, Rosetta, that serves as an oracle and computes a binding score for an input virus-antibody pair. However, score calculation for each possible such pair is intractable. %, as the search space is of the order 10^{130}. We propose a three-pronged approach to address this: first, application of local search, using a native antibody sequence as leverage, second, machine learning to predict binding for antibody-virus pairs, and third, a poisson regression to predict escape costs as a function of antibody sequence assignment. We demonstrate the effectiveness of the proposed methods, and exhibit an antibody with a far higher escape cost (7) than the native (1).