Meiyu Zhong

Meiyu Zhong, Ravi Tandon

As machine learning (ML) based systems are adopted in domains such as law enforcement, criminal justice, finance, hiring and admissions, ensuring the fairness of ML aided decision-making is becoming increasingly important. In this paper, we focus on the problem of fair classification, and introduce a novel min-max F-divergence regularization framework for learning fair classification models while preserving high accuracy. Our framework consists of two trainable networks, namely, a classifier network and a bias/fairness estimator network, where the fairness is measured using the statistical notion of F-divergence. We show that F-divergence measures possess convexity and differentiability properties, and their variational representation make them widely applicable in practical gradient based training methods. The proposed framework can be readily adapted to multiple sensitive attributes and for high dimensional datasets. We study the F-divergence based training paradigm for two types of group fairness constraints, namely, demographic parity and equalized odds. We present a comprehensive set of experiments for several real-world data sets arising in multiple domains (including COMPAS, Law Admissions, Adult Income, and CelebA datasets). To quantify the fairness-accuracy tradeoff, we introduce the notion of fairness-accuracy receiver operating characteristic (FA-ROC) and a corresponding \textit{low-bias} FA-ROC, which we argue is an appropriate measure to evaluate different classifiers. In comparison to several existing approaches for learning fair classifiers (including pre-processing, post-processing and other regularization methods), we show that the proposed F-divergence based framework achieves state-of-the-art performance with respect to the trade-off between accuracy and fairness.

Meiyu Zhong, Ravi Tandon

Certifiable robustness gives the guarantee that small perturbations around an input to a classifier will not change the prediction. There are two approaches to provide certifiable robustness to adversarial examples: a) explicitly training classifiers with small Lipschitz constants, and b) Randomized smoothing, which adds random noise to the input to create a smooth classifier. We propose SPLITZ, a practical and novel approach which leverages the synergistic benefits of both the above ideas into a single framework. Our main idea is to split a classifier into two halves, constrain the Lipschitz constant of the first half, and smooth the second half via randomization. Motivation for SPLITZ comes from the observation that many standard deep networks exhibit heterogeneity in Lipschitz constants across layers. SPLITZ can exploit this heterogeneity while inheriting the scalability of randomized smoothing. We present a principled approach to train SPLITZ and provide theoretical analysis to derive certified robustness guarantees during inference. We present a comprehensive comparison of robustness-accuracy trade-offs and show that SPLITZ consistently improves on existing state-of-the-art approaches in the MNIST, CIFAR-10 and ImageNet datasets. For instance, with $\ell_2$ norm perturbation budget of $ε=1$, SPLITZ achieves $43.2\%$ top-1 test accuracy on CIFAR-10 dataset compared to state-of-art top-1 test accuracy $39.8\%$.

Meiyu Zhong, Ravi Tandon

With the growing adoption of machine learning (ML) systems in areas like law enforcement, criminal justice, finance, hiring, and admissions, it is increasingly critical to guarantee the fairness of decisions assisted by ML. In this paper, we study the tradeoff between fairness and accuracy under the statistical notion of equalized odds. We present a new upper bound on the accuracy (that holds for any classifier), as a function of the fairness budget. In addition, our bounds also exhibit dependence on the underlying statistics of the data, labels and the sensitive group attributes. We validate our theoretical upper bounds through empirical analysis on three real-world datasets: COMPAS, Adult, and Law School. Specifically, we compare our upper bound to the tradeoffs that are achieved by various existing fair classifiers in the literature. Our results show that achieving high accuracy subject to a low-bias could be fundamentally limited based on the statistical disparity across the groups.

Meiyu Zhong, Noel Teku, Ravi Tandon

Speculative Decoding (SD) is a recently proposed technique for faster inference using Large Language Models (LLMs). SD operates by using a smaller draft LLM for autoregressively generating a sequence of tokens and a larger target LLM for parallel verification to ensure statistical consistency. However, periodic parallel calls to the target LLM for verification prevent SD from achieving even lower latencies. We propose SPRINTER, which utilizes a low-complexity verifier trained to predict if tokens generated from a draft LLM would be accepted by the target LLM. By performing sequential approximate verification, SPRINTER does not require verification by the target LLM and is only invoked when a token is deemed unacceptable. This reduces the number of calls to the larger LLM, achieving further speedups and lower computation cost. We present a theoretical analysis of SPRINTER, examining the statistical properties of the generated tokens, as well as the expected reduction in latency as a function of the verifier. We evaluate SPRINTER on several datasets and model pairs, demonstrating that approximate verification can still maintain high quality generation while further reducing latency.

Meiyu Zhong, Ravi Tandon

Adversarial training is one of the predominant techniques for training classifiers that are robust to adversarial attacks. Recent work, however has found that adversarial training, which makes the overall classifier robust, it does not necessarily provide equal amount of robustness for all classes. In this paper, we propose the use of mixup for the problem of learning fair robust classifiers, which can provide similar robustness across all classes. Specifically, the idea is to mix inputs from the same classes and perform adversarial training on mixed up inputs. We present a theoretical analysis of this idea for the case of linear classifiers and show that mixup combined with adversarial training can provably reduce the class-wise robustness disparity. This method not only contributes to reducing the disparity in class-wise adversarial risk, but also the class-wise natural risk. Complementing our theoretical analysis, we also provide experimental results on both synthetic data and the real world dataset (CIFAR-10), which shows improvement in class wise disparities for both natural and adversarial risks.