Vivek Narayanaswamy

Jayaraman J. Thiagarajan, Rushil Anirudh, Vivek Narayanaswamy et al.

We are interested in estimating the uncertainties of deep neural networks, which play an important role in many scientific and engineering problems. In this paper, we present a striking new finding that an ensemble of neural networks with the same weight initialization, trained on datasets that are shifted by a constant bias gives rise to slightly inconsistent trained models, where the differences in predictions are a strong indicator of epistemic uncertainties. Using the neural tangent kernel (NTK), we demonstrate that this phenomena occurs in part because the NTK is not shift-invariant. Since this is achieved via a trivial input transformation, we show that this behavior can therefore be approximated by training a single neural network -- using a technique that we call $Δ-$UQ -- that estimates uncertainty around prediction by marginalizing out the effect of the biases during inference. We show that $Δ-$UQ's uncertainty estimates are superior to many of the current methods on a variety of benchmarks -- outlier rejection, calibration under distribution shift, and sequential design optimization of black box functions. Code for $Δ-$UQ can be accessed at https://github.com/LLNL/DeltaUQ

Grace Billingsley, Julia Dietlmeier, Vivek Narayanaswamy et al.

We propose an accurate and fast classification network for classification of brain tumors in MRI images that outperforms all lightweight methods investigated in terms of accuracy. We test our model on a challenging 2D T1-weighted CE-MRI dataset containing three types of brain tumors: Meningioma, Glioma and Pituitary. We introduce an l2-normalized spatial attention mechanism that acts as a regularizer against overfitting during training. We compare our results against the state-of-the-art on this dataset and show that by integrating l2-normalized spatial attention into a baseline network we achieve a performance gain of 1.79 percentage points. Even better accuracy can be attained by combining our model in an ensemble with the pretrained VGG16 at the expense of execution speed. Our code is publicly available at https://github.com/juliadietlmeier/MRI_image_classification

Rakshith Subramanyam, Kowshik Thopalli, Vivek Narayanaswamy et al.

Reliably detecting when a deployed machine learning model is likely to fail on a given input is crucial for ensuring safe operation. In this work, we propose DECIDER (Debiasing Classifiers to Identify Errors Reliably), a novel approach that leverages priors from large language models (LLMs) and vision-language models (VLMs) to detect failures in image classification models. DECIDER utilizes LLMs to specify task-relevant core attributes and constructs a ``debiased'' version of the classifier by aligning its visual features to these core attributes using a VLM, and detects potential failure by measuring disagreement between the original and debiased models. In addition to proactively identifying samples on which the model would fail, DECIDER also provides human-interpretable explanations for failure through a novel attribute-ablation strategy. Through extensive experiments across diverse benchmarks spanning subpopulation shifts (spurious correlations, class imbalance) and covariate shifts (synthetic corruptions, domain shifts), DECIDER consistently achieves state-of-the-art failure detection performance, significantly outperforming baselines in terms of the overall Matthews correlation coefficient as well as failure and success recall. Our codes can be accessed at~\url{https://github.com/kowshikthopalli/DECIDER/}

Vivek Narayanaswamy, Yamen Mubarka, Rushil Anirudh et al.

We focus on the problem of producing well-calibrated out-of-distribution (OOD) detectors, in order to enable safe deployment of medical image classifiers. Motivated by the difficulty of curating suitable calibration datasets, synthetic augmentations have become highly prevalent for inlier/outlier specification. While there have been rapid advances in data augmentation techniques, this paper makes a striking finding that the space in which the inliers and outliers are synthesized, in addition to the type of augmentation, plays a critical role in calibrating OOD detectors. Using the popular energy-based OOD detection framework, we find that the optimal protocol is to synthesize latent-space inliers along with diverse pixel-space outliers. Based on empirical studies with multiple medical imaging benchmarks, we demonstrate that our approach consistently leads to superior OOD detection ($15\% - 35\%$ in AUROC) over the state-of-the-art in a variety of open-set recognition settings.

Aditya Taparia, Ransalu Senanayake, Kowshik Thopalli et al.

Understanding why a large language model (LLM) is uncertain about the response is important for their reliable deployment. Current approaches, which either provide a single uncertainty score or rely on the classical aleatoric-epistemic dichotomy, fail to offer actionable insights for improving the generative model. Recent studies have also shown that such methods are not enough for understanding uncertainty in LLMs. In this work, we advocate for an uncertainty decomposition framework that dissects LLM uncertainty into three distinct semantic components: (i) input ambiguity, arising from ambiguous prompts; (ii) knowledge gaps, caused by insufficient parametric evidence; and (iii) decoding randomness, stemming from stochastic sampling. Through a series of experiments we demonstrate that the dominance of these components can shift across model size and task. Our framework provides a better understanding to audit LLM reliability and detect hallucinations, paving the way for targeted interventions and more trustworthy systems.

Akshay Kulkarni, Tsui-Wei Weng, Vivek Narayanaswamy et al.

Sparse autoencoders (SAEs) promise a unified approach for mechanistic interpretability, concept discovery, and model steering in LLMs and LVLMs. However, realizing this potential requires that the learned features be both interpretable and steerable. To that end, we introduce two new computationally inexpensive interpretability and steerability metrics and conduct a systematic analysis on LVLMs. Our analysis uncovers two observations; (i) a majority of SAE neurons exhibit either low interpretability or low steerability or both, rendering them ineffective for downstream use; and (ii) due to the unsupervised nature of SAEs, user-desired concepts are often absent in the learned dictionary, thus limiting their practical utility. To address these limitations, we propose Concept Bottleneck Sparse Autoencoders (CB-SAE) - a novel post-hoc framework that prunes low-utility neurons and augments the latent space with a lightweight concept bottleneck aligned to a user-defined concept set. The resulting CB-SAE improves interpretability by +32.1% and steerability by +14.5% across LVLMs and image generation tasks. We will make our code and model weights available.

Jayaraman J. Thiagarajan, Vivek Narayanaswamy, Puja Trivedi et al.

Safe deployment of AI models requires proactive detection of failures to prevent costly errors. To this end, we study the important problem of detecting failures in deep regression models. Existing approaches rely on epistemic uncertainty estimates or inconsistency w.r.t the training data to identify failure. Interestingly, we find that while uncertainties are necessary they are insufficient to accurately characterize failure in practice. Hence, we introduce PAGER (Principled Analysis of Generalization Errors in Regressors), a framework to systematically detect and characterize failures in deep regressors. Built upon the principle of anchored training in deep models, PAGER unifies both epistemic uncertainty and complementary manifold non-conformity scores to accurately organize samples into different risk regimes.

Vivek Narayanaswamy, Kowshik Thopalli, Bhavya Kailkhura et al.

Sparse autoencoders (SAEs) are widely used in mechanistic interpretability to project LLM activations onto sparse latent spaces. However, sparsity alone is an imperfect proxy for interpretability, and current training objectives often result in brittle latent representations. SAEs are known to be prone to feature absorption, where general features are subsumed by more specific ones due to co-occurrence, degrading interpretability despite high reconstruction fidelity. Recent negative results on Out-of-Distribution (OOD) performance further underscore broader robustness related failures tied to under-specified training objectives. We address this by proposing a masking-based regularization that randomly replaces tokens during training to disrupt co-occurrence patterns. This improves robustness across SAE architectures and sparsity levels reducing absorption, enhancing probing performance, and narrowing the OOD gap. Our results point toward a practical path for more reliable interpretability tools.

Srikar Yellapragada, Kowshik Thopalli, Vivek Narayanaswamy et al.

Vision Transformers (ViTs) have shown success across a variety of tasks due to their ability to capture global image representations. Recent studies have identified the existence of high-norm tokens in ViTs, which can interfere with unsupervised object discovery. To address this, the use of "registers" which are additional tokens that isolate high norm patch tokens while capturing global image-level information has been proposed. While registers have been studied extensively for object discovery, their generalization properties particularly in out-of-distribution (OOD) scenarios, remains underexplored. In this paper, we examine the utility of register token embeddings in providing additional features for improving generalization and anomaly rejection. To that end, we propose a simple method that combines the special CLS token embedding commonly employed in ViTs with the average-pooled register embeddings to create feature representations which are subsequently used for training a downstream classifier. We find that this enhances OOD generalization and anomaly rejection, while maintaining in-distribution (ID) performance. Extensive experiments across multiple ViT backbones trained with and without registers reveal consistent improvements of 2-4\% in top-1 OOD accuracy and a 2-3\% reduction in false positive rates for anomaly detection. Importantly, these gains are achieved without additional computational overhead.

Vivek Narayanaswamy, Kowshik Thopalli, Rushil Anirudh et al.

Anchoring is a recent, architecture-agnostic principle for training deep neural networks that has been shown to significantly improve uncertainty estimation, calibration, and extrapolation capabilities. In this paper, we systematically explore anchoring as a general protocol for training vision models, providing fundamental insights into its training and inference processes and their implications for generalization and safety. Despite its promise, we identify a critical problem in anchored training that can lead to an increased risk of learning undesirable shortcuts, thereby limiting its generalization capabilities. To address this, we introduce a new anchored training protocol that employs a simple regularizer to mitigate this issue and significantly enhances generalization. We empirically evaluate our proposed approach across datasets and architectures of varying scales and complexities, demonstrating substantial performance gains in generalization and safety metrics compared to the standard training protocol.

Jayaraman J. Thiagarajan, Vivek Narayanaswamy, Deepta Rajan et al.

Explanation techniques that synthesize small, interpretable changes to a given image while producing desired changes in the model prediction have become popular for introspecting black-box models. Commonly referred to as counterfactuals, the synthesized explanations are required to contain discernible changes (for easy interpretability) while also being realistic (consistency to the data manifold). In this paper, we focus on the case where we have access only to the trained deep classifier and not the actual training data. While the problem of inverting deep models to synthesize images from the training distribution has been explored, our goal is to develop a deep inversion approach to generate counterfactual explanations for a given query image. Despite their effectiveness in conditional image synthesis, we show that existing deep inversion methods are insufficient for producing meaningful counterfactuals. We propose DISC (Deep Inversion for Synthesizing Counterfactuals) that improves upon deep inversion by utilizing (a) stronger image priors, (b) incorporating a novel manifold consistency objective and (c) adopting a progressive optimization strategy. We find that, in addition to producing visually meaningful explanations, the counterfactuals from DISC are effective at learning classifier decision boundaries and are robust to unknown test-time corruptions.

Vivek Narayanaswamy, Jayaraman J. Thiagarajan, Deepta Rajan et al.

With increased interest in adopting AI methods for clinical diagnosis, a vital step towards safe deployment of such tools is to ensure that the models not only produce accurate predictions but also do not generalize to data regimes where the training data provide no meaningful evidence. Existing approaches for ensuring the distribution of model predictions to be similar to that of the true distribution rely on explicit uncertainty estimators that are inherently hard to calibrate. In this paper, we propose to train a loss estimator alongside the predictive model, using a contrastive training objective, to directly estimate the prediction uncertainties. Interestingly, we find that, in addition to producing well-calibrated uncertainties, this approach improves the generalization behavior of the predictor. Using a dermatology use-case, we show the impact of loss estimators on model generalization, in terms of both its fidelity on in-distribution data and its ability to detect out of distribution samples or new classes unseen during training.

Vivek Narayanaswamy, Jayaraman J. Thiagarajan, Andreas Spanias

Through the use of carefully tailored convolutional neural network architectures, a deep image prior (DIP) can be used to obtain pre-images from latent representation encodings. Though DIP inversion has been known to be superior to conventional regularized inversion strategies such as total variation, such an over-parameterized generator is able to effectively reconstruct even images that are not in the original data distribution. This limitation makes it challenging to utilize such priors for tasks such as counterfactual reasoning, wherein the goal is to generate small, interpretable changes to an image that systematically leads to changes in the model prediction. To this end, we propose a novel regularization strategy based on an auxiliary loss estimator jointly trained with the predictor, which efficiently guides the prior to recover natural pre-images. Our empirical studies with a real-world ISIC skin lesion detection problem clearly evidence the effectiveness of the proposed approach in synthesizing meaningful counterfactuals. In comparison, we find that the standard DIP inversion often proposes visually imperceptible perturbations to irrelevant parts of the image, thus providing no additional insights into the model behavior.

Jayaraman J. Thiagarajan, Vivek Narayanaswamy, Rushil Anirudh et al.

With increasing reliance on the outcomes of black-box models in critical applications, post-hoc explainability tools that do not require access to the model internals are often used to enable humans understand and trust these models. In particular, we focus on the class of methods that can reveal the influence of input features on the predicted outputs. Despite their wide-spread adoption, existing methods are known to suffer from one or more of the following challenges: computational complexities, large uncertainties and most importantly, inability to handle real-world domain shifts. In this paper, we propose PRoFILE, a novel feature importance estimation method that addresses all these challenges. Through the use of a loss estimator jointly trained with the predictive model and a causal objective, PRoFILE can accurately estimate the feature importance scores even under complex distribution shifts, without any additional re-training. To this end, we also develop learning strategies for training the loss estimator, namely contrastive and dropout calibration, and find that it can effectively detect distribution shifts. Using empirical studies on several benchmark image and non-image data, we show significant improvements over state-of-the-art approaches, both in terms of fidelity and robustness.

Vivek Narayanaswamy, Jayaraman J. Thiagarajan, Rushil Anirudh et al.

State-of-the-art under-determined audio source separation systems rely on supervised end-end training of carefully tailored neural network architectures operating either in the time or the spectral domain. However, these methods are severely challenged in terms of requiring access to expensive source level labeled data and being specific to a given set of sources and the mixing process, which demands complete re-training when those assumptions change. This strongly emphasizes the need for unsupervised methods that can leverage the recent advances in data-driven modeling, and compensate for the lack of labeled data through meaningful priors. To this end, we propose a novel approach for audio source separation based on generative priors trained on individual sources. Through the use of projected gradient descent optimization, our approach simultaneously searches in the source-specific latent spaces to effectively recover the constituent sources. Though the generative priors can be defined in the time domain directly, e.g. WaveGAN, we find that using spectral domain loss functions for our optimization leads to good-quality source estimates. Our empirical studies on standard spoken digit and instrument datasets clearly demonstrate the effectiveness of our approach over classical as well as state-of-the-art unsupervised baselines.