Recent neural models that extend the pretrain-then-finetune paradigm continue to achieve new state-of-the-art results on joint goal accuracy (JGA) for dialogue state tracking (DST) benchmarks. However, we call into question their robustness as they show sharp drops in JGA for conversations containing utterances or dialog flows with realistic perturbations. Inspired by CheckList (Ribeiro et al., 2020), we design a collection of metrics called CheckDST that facilitate comparisons of DST models on comprehensive dimensions of robustness by testing well-known weaknesses with augmented test sets. We evaluate recent DST models with CheckDST and argue that models should be assessed more holistically rather than pursuing state-of-the-art on JGA since a higher JGA does not guarantee better overall robustness. We find that span-based classification models are resilient to unseen named entities but not robust to language variety, whereas those based on autoregressive language models generalize better to language variety but tend to memorize named entities and often hallucinate. Due to their respective weaknesses, neither approach is yet suitable for real-world deployment. We believe CheckDST is a useful guide for future research to develop task-oriented dialogue models that embody the strengths of various methods.
We derive information-theoretic lower bounds on the Bayes risk and generalization error of realizable machine learning models. In particular, we employ an analysis in which the rate-distortion function of the model parameters bounds the required mutual information between the training samples and the model parameters in order to learn a model up to a Bayes risk constraint. For realizable models, we show that both the rate distortion function and mutual information admit expressions that are convenient for analysis. For models that are (roughly) lower Lipschitz in their parameters, we bound the rate distortion function from below, whereas for VC classes, the mutual information is bounded above by $d_\mathrm{vc}\log(n)$. When these conditions match, the Bayes risk with respect to the zero-one loss scales no faster than $\Omega(d_\mathrm{vc}/n)$, which matches known outer bounds and minimax lower bounds up to logarithmic factors. We also consider the impact of label noise, providing lower bounds when training and/or test samples are corrupted.
Federated learning (FL) is an emerging privacy-preserving paradigm, where a global model is trained at a central server while keeping client data local. However, FL can still indirectly leak private client information through model updates during training. Differential privacy (DP) can be employed to provide privacy guarantees within FL, typically at the cost of degraded final trained model. In this work, we consider a heterogeneous DP setup where clients are considered private by default, but some might choose to opt out of DP. We propose a new algorithm for federated learning with opt-out DP, referred to as \emph{FeO2}, along with a discussion on its advantages compared to the baselines of private and personalized FL algorithms. We prove that the server-side and client-side procedures in \emph{FeO2} are optimal for a simplified linear problem. We also analyze the incentive for opting out of DP in terms of performance gain. Through numerical experiments, we show that \emph{FeO2} provides up to $9.27\%$ performance gain in the global model compared to the baseline DP FL for the considered datasets. Additionally, we show a gap in the average performance of personalized models between non-private and private clients of up to $3.49\%$, empirically illustrating an incentive for clients to opt out.
While deep learning through empirical risk minimization (ERM) has succeeded at achieving human-level performance at a variety of complex tasks, ERM generalizes poorly to distribution shift. This is partly explained by overfitting to spurious features such as background in images or named entities in natural language. Synthetic data augmentation followed by empirical risk minimization (DA-ERM) is a simple yet powerful solution to remedy this problem. In this paper, we propose data augmented invariant regularization (DAIR). The idea of DAIR is based on the observation that the model performance (loss) is desired to be consistent on the augmented sample and the original one. DAIR introduces a regularizer on DA-ERM to penalize such loss inconsistency. Both theoretically and through empirical experiments, we show that a particular form of the DAIR regularizer consistently performs well in a variety of settings. We apply it to multiple real-world learning problems involving domain shift, namely robust regression, visual question answering, robust deep neural network training, and task-oriented dialog modeling. Our experiments show that DAIR consistently outperforms ERM and DA-ERM with little marginal cost and setting new state-of-the-art results in several benchmarks.
Exponential tilting is a technique commonly used in fields such as statistics, probability, information theory, and optimization to create parametric distribution shifts. Despite its prevalence in related fields, tilting has not seen widespread use in machine learning. In this work, we aim to bridge this gap by exploring the use of tilting in risk minimization. We study a simple extension to ERM -- tilted empirical risk minimization (TERM) -- which uses exponential tilting to flexibly tune the impact of individual losses. The resulting framework has several useful properties: We show that TERM can increase or decrease the influence of outliers, respectively, to enable fairness or robustness; has variance-reduction properties that can benefit generalization; and can be viewed as a smooth approximation to a superquantile method. Our work makes rigorous connections between TERM and related objectives, such as Value-at-Risk, Conditional Value-at-Risk, and distributionally robust optimization (DRO). We develop batch and stochastic first-order optimization methods for solving TERM, provide convergence guarantees for the solvers, and show that the framework can be efficiently solved relative to common alternatives. Finally, we demonstrate that TERM can be used for a multitude of applications in machine learning, such as enforcing fairness between subgroups, mitigating the effect of outliers, and handling class imbalance. Despite the straightforward modification TERM makes to traditional ERM objectives, we find that the framework can consistently outperform ERM and deliver competitive performance with state-of-the-art, problem-specific approaches.
MultiWOZ is one of the most popular multi-domain task-oriented dialog datasets, containing 10K+ annotated dialogs covering eight domains. It has been widely accepted as a benchmark for various dialog tasks, e.g., dialog state tracking (DST), natural language generation (NLG), and end-to-end (E2E) dialog modeling. In this work, we identify an overlooked issue with dialog state annotation inconsistencies in the dataset, where a slot type is tagged inconsistently across similar dialogs leading to confusion for DST modeling. We propose an automated correction for this issue, which is present in a whopping 70% of the dialogs. Additionally, we notice that there is significant entity bias in the dataset (e.g., "cambridge" appears in 50% of the destination cities in the train domain). The entity bias can potentially lead to named entity memorization in generative models, which may go unnoticed as the test set suffers from a similar entity bias as well. We release a new test set with all entities replaced with unseen entities. Finally, we benchmark joint goal accuracy (JGA) of the state-of-the-art DST baselines on these modified versions of the data. Our experiments show that the annotation inconsistency corrections lead to 7-10% improvement in JGA. On the other hand, we observe a 29% drop in JGA when models are evaluated on the new test set with unseen entities.
In this paper, we propose a new notion of fairness violation, called Exponential R\'enyi Mutual Information (ERMI). We show that ERMI is a strong fairness violation notion in the sense that it provides upper bound guarantees on existing notions of fairness violation. We then propose the Fair Empirical Risk Minimization via ERMI regularization framework, called FERMI. Whereas most existing in-processing fairness algorithms are deterministic, we provide the first stochastic optimization method with a provable convergence guarantee for solving FERMI. Our stochastic algorithm is amenable to large-scale problems, as we demonstrate experimentally. In addition, we provide a batch (deterministic) algorithm for solving FERMI with the optimal rate of convergence. Both of our algorithms are applicable to problems with multiple (non-binary) sensitive attributes and non-binary targets. Extensive experiments show that FERMI achieves the most favorable tradeoffs between fairness violation and test accuracy across various problem setups compared with state-of-the-art baselines.
A video-grounded dialogue system is required to understand both dialogue, which contains semantic dependencies from turn to turn, and video, which contains visual cues of spatial and temporal scene variations. Building such dialogue systems is a challenging problem involving complex multimodal and temporal inputs, and studying them independently is hard with existing datasets. Existing benchmarks do not have enough annotations to help analyze dialogue systems and understand their linguistic and visual reasoning capability and limitations in isolation. These benchmarks are also not explicitly designed to minimize biases that models can exploit without actual reasoning. To address these limitations, in this paper, we present a diagnostic dataset that can test a range of reasoning abilities on videos and dialogues. The dataset is designed to contain minimal biases and has detailed annotations for the different types of reasoning each question requires, including cross-turn video interval tracking and dialogue object tracking. We use our dataset to analyze several dialogue system approaches, providing interesting insights into their abilities and limitations. In total, the dataset contains $10$ instances of $10$-round dialogues for each of $\sim11k$ synthetic videos, resulting in more than $100k$ dialogues and $1M$ question-answer pairs. Our code and dataset will be made public.
Models trained in machine learning processes may store information about individual samples used in the training process. There are many cases where the impact of an individual sample may need to be deleted and unlearned (i.e., removed) from the model. Retraining the model from scratch after removing a sample from its training set guarantees perfect unlearning, however, it becomes increasingly expensive as the size of training dataset increases. One solution to this issue is utilizing an ensemble learning method that splits the dataset into disjoint shards and assigns them to non-communicating weak learners and then aggregates their models using a pre-defined rule. This framework introduces a trade-off between performance and unlearning cost which may result in an unreasonable performance degradation, especially as the number of shards increases. In this paper, we present a coded learning protocol where the dataset is linearly coded before the learning phase. We also present the corresponding unlearning protocol for the aforementioned coded learning model along with a discussion on the proposed protocol's success in ensuring perfect unlearning. Finally, experimental results show the effectiveness of the coded machine unlearning protocol in terms of performance versus unlearning cost trade-off.
In addition to accuracy, fairness and robustness are two critical concerns for federated learning systems. In this work, we first identify that robustness to adversarial training-time attacks and fairness, measured as the uniformity of performance across devices, are competing constraints in statistically heterogeneous networks. To address these constraints, we propose employing a simple, general multi-task learning objective, and analyze the ability of the objective to achieve a favorable tradeoff between fairness and robustness. We develop a scalable solver for the objective and show that multi-task learning can enable more accurate, robust, and fair models relative to state-of-the-art baselines across a suite of federated datasets.