Akshay Mehra

Akshay Mehra, Bhavya Kailkhura, Pin-Yu Chen et al.

Domain Generalization (DG) aims to learn models whose performance remains high on unseen domains encountered at test-time by using data from multiple related source domains. Many existing DG algorithms reduce the divergence between source distributions in a representation space to potentially align the unseen domain close to the sources. This is motivated by the analysis that explains generalization to unseen domains using distributional distance (such as the Wasserstein distance) to the sources. However, due to the openness of the DG objective, it is challenging to evaluate DG algorithms comprehensively using a few benchmark datasets. In particular, we demonstrate that the accuracy of the models trained with DG methods varies significantly across unseen domains, generated from popular benchmark datasets. This highlights that the performance of DG methods on a few benchmark datasets may not be representative of their performance on unseen domains in the wild. To overcome this roadblock, we propose a universal certification framework based on distributionally robust optimization (DRO) that can efficiently certify the worst-case performance of any DG method. This enables a data-independent evaluation of a DG method complementary to the empirical evaluations on benchmark datasets. Furthermore, we propose a training algorithm that can be used with any DG method to provably improve their certified performance. Our empirical evaluation demonstrates the effectiveness of our method at significantly improving the worst-case loss (i.e., reducing the risk of failure of these models in the wild) without incurring a significant performance drop on benchmark datasets.

Akshay Mehra, Skyler Seto, Navdeep Jaitly et al.

Multi-Exit models (MEMs) use an early-exit strategy to improve the accuracy and efficiency of deep neural networks (DNNs) by allowing samples to exit the network before the last layer. However, the effectiveness of MEMs in the presence of distribution shifts remains largely unexplored. Our work examines how distribution shifts generated by common image corruptions affect the accuracy/efficiency of MEMs. We find that under common corruptions, early-exiting at the first correct exit reduces the inference cost and provides a significant boost in accuracy ( 10%) over exiting at the last layer. However, with realistic early-exit strategies, which do not assume knowledge about the correct exits, MEMs still reduce inference cost but provide a marginal improvement in accuracy (1%) compared to exiting at the last layer. Moreover, the presence of distribution shift widens the gap between an MEM's maximum classification accuracy and realistic early-exit strategies by 5% on average compared with the gap on in-distribution data. Our empirical analysis shows that the lack of calibration due to a distribution shift increases the susceptibility of such early-exit strategies to exit early and increases misclassification rates. Furthermore, the lack of calibration increases the inconsistency in the predictions of the model across exits, leading to both inefficient inference and more misclassifications compared with evaluation on in-distribution data. Finally, we propose two metrics, underthinking and overthinking, that quantify the different behavior of practical early-exit strategy under distribution shifts, and provide insights into improving the practical utility of MEMs.

Akshay Mehra, Yunbei Zhang, Jihun Hamm

The growing popularity of transfer learning, due to the availability of models pre-trained on vast amounts of data, makes it imperative to understand when the knowledge of these pre-trained models can be transferred to obtain high-performing models on downstream target tasks. However, the exact conditions under which transfer learning succeeds in a cross-domain cross-task setting are still poorly understood. To bridge this gap, we propose a novel analysis that analyzes the transferability of the representations of pre-trained models to downstream tasks in terms of their relatedness to a given reference task. Our analysis leads to an upper bound on transferability in terms of task-relatedness, quantified using the difference between the class priors, label sets, and features of the two tasks. Our experiments using state-of-the-art pre-trained models show the effectiveness of task-relatedness in explaining transferability on various vision and language tasks. The efficient computability of task-relatedness even without labels of the target task and its high correlation with the model's accuracy after end-to-end fine-tuning on the target task makes it a useful metric for transferability estimation. Our empirical results of using task-relatedness to select the best pre-trained model from a model zoo for a target task highlight its utility for practical problems.

Akshay Mehra, Yunbei Zhang, Jihun Hamm

Gauging the performance of ML models on data from unseen domains at test-time is essential yet a challenging problem due to the lack of labels in this setting. Moreover, the performance of these models on in-distribution data is a poor indicator of their performance on data from unseen domains. Thus, it is essential to develop metrics that can provide insights into the model's performance at test time and can be computed only with the information available at test time (such as their model parameters, the training data or its statistics, and the unlabeled test data). To this end, we propose a metric based on Optimal Transport that is highly correlated with the model's performance on unseen domains and is efficiently computable only using information available at test time. Concretely, our metric characterizes the model's performance on unseen domains using only a small amount of unlabeled data from these domains and data or statistics from the training (source) domain(s). Through extensive empirical evaluation using standard benchmark datasets, and their corruptions, we demonstrate the utility of our metric in estimating the model's performance in various practical applications. These include the problems of selecting the source data and architecture that leads to the best performance on data from an unseen domain and the problem of predicting a deployed model's performance at test time on unseen domains. Our empirical results show that our metric, which uses information from both the source and the unseen domain, is highly correlated with the model's performance, achieving a significantly better correlation than that obtained via the popular prediction entropy-based metric, which is computed solely using the data from the unseen domain.

Akshay Mehra, Trisha Mittal, Subhadra Gopalakrishnan et al.

Coreset Selection (CS) aims to identify a subset of the training dataset that achieves model performance comparable to using the entire dataset. Many state-of-the-art CS methods select coresets using scores whose computation requires training the downstream model on the entire dataset first and recording changes in the model's behavior on samples as it trains (training dynamics). These scores are inefficient to compute and hard to interpret, as they do not indicate whether a sample is difficult to learn in general or only for a specific downstream model. Our work addresses these challenges by proposing a score that computes a sample's difficulty using human-understandable textual attributes (concepts) independent of any downstream model. Specifically, we measure the alignment between a sample's visual features and concept bottlenecks, derived via large language models, by training a linear concept bottleneck layer and computing the sample's difficulty score using it.We then use stratified sampling based on this score to generate a coreset of the dataset.Crucially, our score is efficiently computable without training the downstream model on the full dataset even once, leads to high-performing coresets for various downstream models, and is computable even for an unlabeled dataset. Through experiments on CIFAR-10/100, and ImageNet-1K, we show that our coresets outperform random subsets, even at high pruning rates, and achieve model performance comparable to or better than coresets found by training dynamics-based methods.

Yunbei Zhang, Akshay Mehra, Shuaicheng Niu et al.

Continual Test-Time Adaptation (CTTA) seeks to adapt source pre-trained models to continually changing, unseen target domains. While existing CTTA methods assume structured domain changes with uniform durations, real-world environments often exhibit dynamic patterns where domains recur with varying frequencies and durations. Current approaches, which adapt the same parameters across different domains, struggle in such dynamic conditions-they face convergence issues with brief domain exposures, risk forgetting previously learned knowledge, or misapplying it to irrelevant domains. To remedy this, we propose DPCore, a method designed for robust performance across diverse domain change patterns while ensuring computational efficiency. DPCore integrates three key components: Visual Prompt Adaptation for efficient domain alignment, a Prompt Coreset for knowledge preservation, and a Dynamic Update mechanism that intelligently adjusts existing prompts for similar domains while creating new ones for substantially different domains. Extensive experiments on four benchmarks demonstrate that DPCore consistently outperforms various CTTA methods, achieving state-of-the-art performance in both structured and dynamic settings while reducing trainable parameters by 99% and computation time by 64% compared to previous approaches.

Yunbei Zhang, Akshay Mehra, Jihun Hamm

Vision Transformers (ViTs) have demonstrated remarkable capabilities in learning representations, but their performance is compromised when applied to unseen domains. Previous methods either engage in prompt learning during the training phase or modify model parameters at test time through entropy minimization. The former often overlooks unlabeled target data, while the latter doesn't fully address domain shifts. In this work, our approach, Optimal Transport-guided Test-Time Visual Prompting (OT-VP), handles these problems by leveraging prompt learning at test time to align the target and source domains without accessing the training process or altering pre-trained model parameters. This method involves learning a universal visual prompt for the target domain by optimizing the Optimal Transport distance.OT-VP, with only four learned prompt tokens, exceeds state-of-the-art performance across three stylistic datasets-PACS, VLCS, OfficeHome, and one corrupted dataset ImageNet-C. Additionally, OT-VP operates efficiently, both in terms of memory and computation, and is adaptable for extension to online settings.

Akshay Mehra, Bhavya Kailkhura, Pin-Yu Chen et al.

Unsupervised domain adaptation (UDA) enables cross-domain learning without target domain labels by transferring knowledge from a labeled source domain whose distribution differs from that of the target. However, UDA is not always successful and several accounts of `negative transfer' have been reported in the literature. In this work, we prove a simple lower bound on the target domain error that complements the existing upper bound. Our bound shows the insufficiency of minimizing source domain error and marginal distribution mismatch for a guaranteed reduction in the target domain error, due to the possible increase of induced labeling function mismatch. This insufficiency is further illustrated through simple distributions for which the same UDA approach succeeds, fails, and may succeed or fail with an equal chance. Motivated from this, we propose novel data poisoning attacks to fool UDA methods into learning representations that produce large target domain errors. We evaluate the effect of these attacks on popular UDA methods using benchmark datasets where they have been previously shown to be successful. Our results show that poisoning can significantly decrease the target domain accuracy, dropping it to almost 0% in some cases, with the addition of only 10% poisoned data in the source domain. The failure of these UDA methods demonstrates their limitations at guaranteeing cross-domain generalization consistent with our lower bound. Thus, evaluating UDA methods in adversarial settings such as data poisoning provides a better sense of their robustness to data distributions unfavorable for UDA.

Byunggill Joe, Akshay Mehra, Insik Shin et al.

Electronic Health Records (EHRs) provide a wealth of information for machine learning algorithms to predict the patient outcome from the data including diagnostic information, vital signals, lab tests, drug administration, and demographic information. Machine learning models can be built, for example, to evaluate patients based on their predicted mortality or morbidity and to predict required resources for efficient resource management in hospitals. In this paper, we demonstrate that an attacker can manipulate the machine learning predictions with EHRs easily and selectively at test time by backdoor attacks with the poisoned training data. Furthermore, the poison we create has statistically similar features to the original data making it hard to detect, and can also attack multiple machine learning models without any knowledge of the models. With less than 5% of the raw EHR data poisoned, we achieve average attack success rates of 97% on mortality prediction tasks with MIMIC-III database against Logistic Regression, Multilayer Perceptron, and Long Short-term Memory models simultaneously.