Mohit Prabhushankar

Mohit Prabhushankar, Kiran Kokilepersaud, Yash-yee Logan et al.

Clinical diagnosis of the eye is performed over multifarious data modalities including scalar clinical labels, vectorized biomarkers, two-dimensional fundus images, and three-dimensional Optical Coherence Tomography (OCT) scans. Clinical practitioners use all available data modalities for diagnosing and treating eye diseases like Diabetic Retinopathy (DR) or Diabetic Macular Edema (DME). Enabling usage of machine learning algorithms within the ophthalmic medical domain requires research into the relationships and interactions between all relevant data over a treatment period. Existing datasets are limited in that they neither provide data nor consider the explicit relationship modeling between the data modalities. In this paper, we introduce the Ophthalmic Labels for Investigating Visual Eye Semantics (OLIVES) dataset that addresses the above limitation. This is the first OCT and near-IR fundus dataset that includes clinical labels, biomarker labels, disease labels, and time-series patient treatment information from associated clinical trials. The dataset consists of 1268 near-IR fundus images each with at least 49 OCT scans, and 16 biomarkers, along with 4 clinical labels and a disease diagnosis of DR or DME. In total, there are 96 eyes' data averaged over a period of at least two years with each eye treated for an average of 66 weeks and 7 injections. We benchmark the utility of OLIVES dataset for ophthalmic data as well as provide benchmarks and concrete research directions for core and emerging machine learning paradigms within medical image analysis.

Seulgi Kim, Kiran Kokilepersaud, Mohit Prabhushankar et al.

Multi-modal fusion methods often suffer from two types of representation collapse: feature collapse where individual dimensions lose their discriminative power (as measured by eigenspectra), and modality collapse where one dominant modality overwhelms the other. Applications like human action anticipation that require fusing multifarious sensor data are hindered by both feature and modality collapse. However, existing methods attempt to counter feature collapse and modality collapse separately. This is because there is no unifying framework that efficiently addresses feature and modality collapse in conjunction. In this paper, we posit the utility of effective rank as an informative measure that can be utilized to quantify and counter both the representation collapses. We propose \textit{Rank-enhancing Token Fuser}, a theoretically grounded fusion framework that selectively blends less informative features from one modality with complementary features from another modality. We show that our method increases the effective rank of the fused representation. To address modality collapse, we evaluate modality combinations that mutually increase each others' effective rank. We show that depth maintains representational balance when fused with RGB, avoiding modality collapse. We validate our method on action anticipation, where we present \texttt{R3D}, a depth-informed fusion framework. Extensive experiments on NTURGBD, UTKinect, and DARai demonstrate that our approach significantly outperforms prior state-of-the-art methods by up to 3.74\%. Our code is available at: \href{https://github.com/olivesgatech/R3D}{https://github.com/olivesgatech/R3D}.

William Stevens, Mohit Prabhushankar, Ghassan AlRegib

Epistemic uncertainty estimation is essential for identifying regions where deep learning system outputs may be unreliable. However, existing approaches require computationally expensive ensemble methods or multiple stochastic forward passes, limiting their scalability to dense prediction tasks like segmentation. We propose Resolution-Aggregated Decoder Mutual Information (RADMI), a single-pass method that estimates prediction uncertainty by measuring mutual information (MI) between consecutive decoder layers in segmentation networks. We observe that elevated inter-layer MI correlates with prediction uncertainty, as the network must integrate conflicting contextual information at ambiguous regions such as class boundaries. Evaluating on a seismic facies segmentation benchmark, RADMI achieves the highest correlation with deep ensemble uncertainty among all single-pass methods, outperforming the next-best baselines by 5.5% in Pearson and 10.7% in Spearman correlation coefficients. Compared to baselines that either lack spatial precision or demand significant computational overhead, RADMI yields sharp, boundary-localized uncertainty maps without architectural modifications. Our results suggest that linear aggregation of normalized information flow provides a principled and efficient proxy for prediction uncertainty in encoder-decoder architectures.

Mohit Prabhushankar, Ghassan AlRegib

In this paper, we advocate for two stages in a neural network's decision making process. The first is the existing feed-forward inference framework where patterns in given data are sensed and associated with previously learned patterns. The second stage is a slower reflection stage where we ask the network to reflect on its feed-forward decision by considering and evaluating all available choices. Together, we term the two stages as introspective learning. We use gradients of trained neural networks as a measurement of this reflection. A simple three-layered Multi Layer Perceptron is used as the second stage that predicts based on all extracted gradient features. We perceptually visualize the post-hoc explanations from both stages to provide a visual grounding to introspection. For the application of recognition, we show that an introspective network is 4% more robust and 42% less prone to calibration errors when generalizing to noisy data. We also illustrate the value of introspective networks in downstream tasks that require generalizability and calibration including active learning, out-of-distribution detection, and uncertainty estimation. Finally, we ground the proposed machine introspection to human introspection for the application of image quality assessment.

Kiran Kokilepersaud, Mohit Prabhushankar, Ghassan AlRegib

In seismic interpretation, pixel-level labels of various rock structures can be time-consuming and expensive to obtain. As a result, there oftentimes exists a non-trivial quantity of unlabeled data that is left unused simply because traditional deep learning methods rely on access to fully labeled volumes. To rectify this problem, contrastive learning approaches have been proposed that use a self-supervised methodology in order to learn useful representations from unlabeled data. However, traditional contrastive learning approaches are based on assumptions from the domain of natural images that do not make use of seismic context. In order to incorporate this context within contrastive learning, we propose a novel positive pair selection strategy based on the position of slices within a seismic volume. We show that the learnt representations from our method out-perform a state of the art contrastive learning methodology in a semantic segmentation task.

Jinsol Lee, Mohit Prabhushankar, Ghassan AlRegib

We propose to utilize gradients for detecting adversarial and out-of-distribution samples. We introduce confounding labels -- labels that differ from normal labels seen during training -- in gradient generation to probe the effective expressivity of neural networks. Gradients depict the amount of change required for a model to properly represent given inputs, providing insight into the representational power of the model established by network architectural properties as well as training data. By introducing a label of different design, we remove the dependency on ground truth labels for gradient generation during inference. We show that our gradient-based approach allows for capturing the anomaly in inputs based on the effective expressivity of the models with no hyperparameter tuning or additional processing, and outperforms state-of-the-art methods for adversarial and out-of-distribution detection.

Ryan Benkert, Mohit Prabhushankar, Ghassan AlRegib et al.

In active learning, acquisition functions define informativeness directly on the representation position within the model manifold. However, for most machine learning models (in particular neural networks) this representation is not fixed due to the training pool fluctuations in between active learning rounds. Therefore, several popular strategies are sensitive to experiment parameters (e.g. architecture) and do not consider model robustness to out-of-distribution settings. To alleviate this issue, we propose a grounded second-order definition of information content and sample importance within the context of active learning. Specifically, we define importance by how often a neural network "forgets" a sample during training - artifacts of second order representation shifts. We show that our definition produces highly accurate importance scores even when the model representations are constrained by the lack of training data. Motivated by our analysis, we develop Gaussian Switch Sampling (GauSS). We show that GauSS is setup agnostic and robust to anomalous distributions with exhaustive experiments on three in-distribution benchmarks, three out-of-distribution benchmarks, and three different architectures. We report an improvement of up to 5% when compared against four popular query strategies.

Ryan Benkert, Mohit Prabhushankar, Ghassan AlRegib

This paper considers deep out-of-distribution active learning. In practice, fully trained neural networks interact randomly with out-of-distribution (OOD) inputs and map aberrant samples randomly within the model representation space. Since data representations are direct manifestations of the training distribution, the data selection process plays a crucial role in outlier robustness. For paradigms such as active learning, this is especially challenging since protocols must not only improve performance on the training distribution most effectively but further render a robust representation space. However, existing strategies directly base the data selection on the data representation of the unlabeled data which is random for OOD samples by definition. For this purpose, we introduce forgetful active learning with switch events (FALSE) - a novel active learning protocol for out-of-distribution active learning. Instead of defining sample importance on the data representation directly, we formulate "informativeness" with learning difficulty during training. Specifically, we approximate how often the network "forgets" unlabeled samples and query the most "forgotten" samples for annotation. We report up to 4.5\% accuracy improvements in over 270 experiments, including four commonly used protocols, two OOD benchmarks, one in-distribution benchmark, and three different architectures.

Jinsol Lee, Charlie Lehman, Mohit Prabhushankar et al.

We analyze the data-dependent capacity of neural networks and assess anomalies in inputs from the perspective of networks during inference. The notion of data-dependent capacity allows for analyzing the knowledge base of a model populated by learned features from training data. We define purview as the additional capacity necessary to characterize inference samples that differ from the training data. To probe the purview of a network, we utilize gradients to measure the amount of change required for the model to characterize the given inputs more accurately. To eliminate the dependency on ground-truth labels in generating gradients, we introduce confounding labels that are formulated by combining multiple categorical labels. We demonstrate that our gradient-based approach can effectively differentiate inputs that cannot be accurately represented with learned features. We utilize our approach in applications of detecting anomalous inputs, including out-of-distribution, adversarial, and corrupted samples. Our approach requires no hyperparameter tuning or additional data processing and outperforms state-of-the-art methods by up to 2.7%, 19.8%, and 35.6% of AUROC scores, respectively.

Zoe Fowler, Kiran Kokilepersaud, Mohit Prabhushankar et al.

This paper presents a novel approach to active learning that takes into account the non-independent and identically distributed (non-i.i.d.) structure of a clinical trial setting. There exists two types of clinical trials: retrospective and prospective. Retrospective clinical trials analyze data after treatment has been performed; prospective clinical trials collect data as treatment is ongoing. Typically, active learning approaches assume the dataset is i.i.d. when selecting training samples; however, in the case of clinical trials, treatment results in a dependency between the data collected at the current and past visits. Thus, we propose prospective active learning to overcome the limitations present in traditional active learning methods and apply it to disease detection in optical coherence tomography (OCT) images, where we condition on the time an image was collected to enforce the i.i.d. assumption. We compare our proposed method to the traditional active learning paradigm, which we refer to as retrospective in nature. We demonstrate that prospective active learning outperforms retrospective active learning in two different types of test settings.

Kiran Kokilepersaud, Yash-Yee Logan, Ryan Benkert et al.

In this paper, we introduce the FOCAL (Ford-OLIVES Collaboration on Active Learning) dataset which enables the study of the impact of annotation-cost within a video active learning setting. Annotation-cost refers to the time it takes an annotator to label and quality-assure a given video sequence. A practical motivation for active learning research is to minimize annotation-cost by selectively labeling informative samples that will maximize performance within a given budget constraint. However, previous work in video active learning lacks real-time annotation labels for accurately assessing cost minimization and instead operates under the assumption that annotation-cost scales linearly with the amount of data to annotate. This assumption does not take into account a variety of real-world confounding factors that contribute to a nonlinear cost such as the effect of an assistive labeling tool and the variety of interactions within a scene such as occluded objects, weather, and motion of objects. FOCAL addresses this discrepancy by providing real annotation-cost labels for 126 video sequences across 69 unique city scenes with a variety of weather, lighting, and seasonal conditions. We also introduce a set of conformal active learning algorithms that take advantage of the sequential structure of video data in order to achieve a better trade-off between annotation-cost and performance while also reducing floating point operations (FLOPS) overhead by at least 77.67%. We show how these approaches better reflect how annotations on videos are done in practice through a sequence selection framework. We further demonstrate the advantage of these approaches by introducing two performance-cost metrics and show that the best conformal active learning method is cheaper than the best traditional active learning method by 113 hours.

Kiran Kokilepersaud, Mohit Prabhushankar, Yavuz Yarici et al.

In this work, we present a methodology to shape a fisheye-specific representation space that reflects the interaction between distortion and semantic context present in this data modality. Fisheye data has the wider field of view advantage over other types of cameras, but this comes at the expense of high radial distortion. As a result, objects further from the center exhibit deformations that make it difficult for a model to identify their semantic context. While previous work has attempted architectural and training augmentation changes to alleviate this effect, no work has attempted to guide the model towards learning a representation space that reflects this interaction between distortion and semantic context inherent to fisheye data. We introduce an approach to exploit this relationship by first extracting distortion class labels based on an object's distance from the center of the image. We then shape a backbone's representation space with a weighted contrastive loss that constrains objects of the same semantic class and distortion class to be close to each other within a lower dimensional embedding space. This backbone trained with both semantic and distortion information is then fine-tuned within an object detection setting to empirically evaluate the quality of the learnt representation. We show this method leads to performance improvements by as much as 1.1% mean average precision over standard object detection strategies and .6% improvement over other state of the art representation learning approaches.

Chen Zhou, Mohit Prabhushankar, Ghassan AlRegib

Humans exhibit disagreement during data labeling. We term this disagreement as human label uncertainty. In this work, we study the ramifications of human label uncertainty (HLU). Our evaluation of existing uncertainty estimation algorithms, with the presence of HLU, indicates the limitations of existing uncertainty metrics and algorithms themselves in response to HLU. Meanwhile, we observe undue effects in predictive uncertainty and generalizability. To mitigate the undue effects, we introduce a novel natural scene statistics (NSS) based label dilution training scheme without requiring massive human labels. Specifically, we first select a subset of samples with low perceptual quality ranked by statistical regularities of images. We then assign separate labels to each sample in this subset to obtain a training set with diluted labels. Our experiments and analysis demonstrate that training with NSS-based label dilution alleviates the undue effects caused by HLU.

Kiran Kokilepersaud, Mohit Prabhushankar, Ghassan AlRegib

This paper presents a novel positive and negative set selection strategy for contrastive learning of medical images based on labels that can be extracted from clinical data. In the medical field, there exists a variety of labels for data that serve different purposes at different stages of a diagnostic and treatment process. Clinical labels and biomarker labels are two examples. In general, clinical labels are easier to obtain in larger quantities because they are regularly collected during routine clinical care, while biomarker labels require expert analysis and interpretation to obtain. Within the field of ophthalmology, previous work has shown that clinical values exhibit correlations with biomarker structures that manifest within optical coherence tomography (OCT) scans. We exploit this relationship between clinical and biomarker data to improve performance for biomarker classification. This is accomplished by leveraging the larger amount of clinical data as pseudo-labels for our data without biomarker labels in order to choose positive and negative instances for training a backbone network with a supervised contrastive loss. In this way, a backbone network learns a representation space that aligns with the clinical data distribution available. Afterwards, we fine-tune the network trained in this manner with the smaller amount of biomarker labeled data with a cross-entropy loss in order to classify these key indicators of disease directly from OCT scans. Our method is shown to outperform state of the art self-supervised methods by as much as 5% in terms of accuracy on individual biomarker detection.

Seulgi Kim, Mohit Prabhushankar, Ghassan AlRegib

Vision Language models (VLMs) have demonstrated strong performance across a wide range of benchmarks, yet they often suffer from modality dominance, where predictions rely disproportionately on a single modality. Prior approaches primarily address this issue by steering model's attention allocation, implicitly assuming that all modalities provide sufficient information. However, attention only determines where the model focuses, and cannot enrich information that is missing or ambiguous. In the real world, input modalities often differ in information density and their signal-to-noise ratios. In such cases, simply adjusting model's attention does not resolve the underlying lack of information. In this paper, we propose \textsc{MoIR}: \textit{Multi-modal Information Router}, an information-level fusion method that explicitly reduces information disparity prior to fusion. \textsc{MoIR} identifies less informative tokens and routes complementary information from a stronger modality, constructing information-dense token representations before they are processed by a large language model. By modifying information availability, \textsc{MoIR} enables reliable shifts in modality dominance, even when one modality is degraded. We evaluate \textsc{MoIR} on three widely used multi-modal benchmarks across multiple model backbones. Experimental results show that \textsc{MoIR} consistently demonstrates more balanced modality contribution, and improves robustness and downstream performance, particularly even under modality degradation. These findings demonstrate that explicitly modifying cross-modal information is an effective and complementary strategy for mitigating modality dominance in multi-modal reasoning models.

Mohit Prabhushankar, Ghassan AlRegib

This paper conjectures and validates a framework that allows for action during inference in supervised neural networks. Supervised neural networks are constructed with the objective to maximize their performance metric in any given task. This is done by reducing free energy and its associated surprisal during training. However, the bottom-up inference nature of supervised networks is a passive process that renders them fallible to noise. In this paper, we provide a thorough background of supervised neural networks, both generative and discriminative, and discuss their functionality from the perspective of free energy principle. We then provide a framework for introducing action during inference. We introduce a new measurement called stochastic surprisal that is a function of the network, the input, and any possible action. This action can be any one of the outputs that the neural network has learnt, thereby lending stochasticity to the measurement. Stochastic surprisal is validated on two applications: Image Quality Assessment and Recognition under noisy conditions. We show that, while noise characteristics are ignored to make robust recognition, they are analyzed to estimate image quality scores. We apply stochastic surprisal on two applications, three datasets, and as a plug-in on twelve networks. In all, it provides a statistically significant increase among all measures. We conclude by discussing the implications of the proposed stochastic surprisal in other areas of cognitive psychology including expectancy-mismatch and abductive reasoning.

Mohit Prabhushankar, Kiran Kokilepersaud, Jorge Quesada et al.

Crowdsourcing annotations has created a paradigm shift in the availability of labeled data for machine learning. Availability of large datasets has accelerated progress in common knowledge applications involving visual and language data. However, specialized applications that require expert labels lag in data availability. One such application is fault segmentation in subsurface imaging. Detecting, tracking, and analyzing faults has broad societal implications in predicting fluid flows, earthquakes, and storing excess atmospheric CO$_2$. However, delineating faults with current practices is a labor-intensive activity that requires precise analysis of subsurface imaging data by geophysicists. In this paper, we propose the $\texttt{CRACKS}$ dataset to detect and segment faults in subsurface images by utilizing crowdsourced resources. We leverage Amazon Mechanical Turk to obtain fault delineations from sections of the Netherlands North Sea subsurface images from (i) $26$ novices who have no exposure to subsurface data and were shown a video describing and labeling faults, (ii) $8$ practitioners who have previously interacted and worked on subsurface data, (iii) one geophysicist to label $7636$ faults in the region. Note that all novices, practitioners, and the expert segment faults on the same subsurface volume with disagreements between and among the novices and practitioners. Additionally, each fault annotation is equipped with the confidence level of the annotator. The paper provides benchmarks on detecting and segmenting the expert labels, given the novice and practitioner labels. Additional details along with the dataset links and codes are available at $\href{https://alregib.ece.gatech.edu/cracks-crowdsourcing-resources-for-analysis-and-categorization-of-key-subsurface-faults/}{link}$.

Prithwijit Chowdhury, Ahmad Mustafa, Mohit Prabhushankar et al.

In geophysics, hydrocarbon prospect risking involves assessing the risks associated with hydrocarbon exploration by integrating data from various sources. Machine learning-based classifiers trained on tabular data have been recently used to make faster decisions on these prospects. The lack of transparency in the decision-making processes of such models has led to the emergence of explainable AI (XAI). LIME and SHAP are two such examples of these XAI methods which try to generate explanations of a particular decision by ranking the input features in terms of importance. However, explanations of the same scenario generated by these two different explanation strategies have shown to disagree or be different, particularly for complex data. This is because the definitions of "importance" and "relevance" differ for different explanation strategies. Thus, grounding these ranked features using theoretically backed causal ideas of necessity and sufficiency can prove to be a more reliable and robust way to improve the trustworthiness of the concerned explanation strategies.We propose a unified framework to generate counterfactuals as well as quantify necessity and sufficiency and use these to perform a robustness evaluation of the explanations provided by LIME and SHAP on high dimensional structured prospect risking data. This robustness test gives us deeper insights into the models capabilities to handle erronous data and which XAI module works best in pair with which model for our dataset for hydorcarbon indication.

Prithwijit Chowdhury, Mohit Prabhushankar, Ghassan AlRegib

The Segment Anything Model (SAM) has revolutionized interactive segmentation through spatial prompting. While existing work primarily focuses on automating prompts in various settings, real-world annotation workflows involve iterative refinement where annotators observe model outputs and strategically place prompts to resolve ambiguities. Current pipelines typically rely on the annotator's visual assessment of the predicted mask quality. We postulate that a principled approach for automated interactive prompting is to use a model-derived criterion to identify the most informative region for the next prompt. In this work, we establish active prompting: a spatial active learning approach where locations within images constitute an unlabeled pool and prompts serve as queries to prioritize information-rich regions, increasing the utility of each interaction. We further present BALD-SAM: a principled framework adapting Bayesian Active Learning by Disagreement (BALD) to spatial prompt selection by quantifying epistemic uncertainty. To do so, we freeze the entire model and apply Bayesian uncertainty modeling only to a small learned prediction head, making intractable uncertainty estimation practical for large multi-million parameter foundation models. Across 16 datasets spanning natural, medical, underwater, and seismic domains, BALD-SAM demonstrates strong cross-domain performance, ranking first or second on 14 of 16 benchmarks. We validate these gains through a comprehensive ablation suite covering 3 SAM backbones and 35 Laplace posterior configurations, amounting to 38 distinct ablation settings. Beyond strong average performance, BALD-SAM surpasses human prompting and, in several categories, even oracle prompting, while consistently outperforming one-shot baselines in final segmentation quality, particularly on thin and structurally complex objects.

Jorge Quesada, Zoe Fowler, Mohammad Alotaibi et al.

The remarkable capabilities of the Segment Anything Model (SAM) for tackling image segmentation tasks in an intuitive and interactive manner has sparked interest in the design of effective visual prompts. Such interest has led to the creation of automated point prompt selection strategies, typically motivated from a feature extraction perspective. However, there is still very little understanding of how appropriate these automated visual prompting strategies are, particularly when compared to humans, across diverse image domains. Additionally, the performance benefits of including such automated visual prompting strategies within the finetuning process of SAM also remains unexplored, as does the effect of interpretable factors like distance between the prompt points on segmentation performance. To bridge these gaps, we leverage a recently released visual prompting dataset, PointPrompt, and introduce a number of benchmarking tasks that provide an array of opportunities to improve the understanding of the way human prompts differ from automated ones and what underlying factors make for effective visual prompts. We demonstrate that the resulting segmentation scores obtained by humans are approximately 29% higher than those given by automated strategies and identify potential features that are indicative of prompting performance with $R^2$ scores over 0.5. Additionally, we demonstrate that performance when using automated methods can be improved by up to 68% via a finetuning approach. Overall, our experiments not only showcase the existing gap between human prompts and automated methods, but also highlight potential avenues through which this gap can be leveraged to improve effective visual prompt design. Further details along with the dataset links and codes are available at https://github.com/olivesgatech/PointPrompt

Seulgi Kim, Ghazal Kaviani, Mohit Prabhushankar et al.

Action anticipation, the task of predicting future actions from partially observed videos, is crucial for advancing intelligent systems. Unlike action recognition, which operates on fully observed videos, action anticipation must handle incomplete information. Hence, it requires temporal reasoning, and inherent uncertainty handling. While recent advances have been made, traditional methods often focus solely on visual modalities, neglecting the potential of integrating multiple sources of information. Drawing inspiration from human behavior, we introduce \textit{Multi-level and Multi-modal Action Anticipation (m\&m-Ant)}, a novel multi-modal action anticipation approach that combines both visual and textual cues, while explicitly modeling hierarchical semantic information for more accurate predictions. To address the challenge of inaccurate coarse action labels, we propose a fine-grained label generator paired with a specialized temporal consistency loss function to optimize performance. Extensive experiments on widely used datasets, including Breakfast, 50 Salads, and DARai, demonstrate the effectiveness of our approach, achieving state-of-the-art results with an average anticipation accuracy improvement of 3.08\% over existing methods. This work underscores the potential of multi-modal and hierarchical modeling in advancing action anticipation and establishes a new benchmark for future research in the field. Our code is available at: https://github.com/olivesgatech/mM-ant.

Mohit Prabhushankar, Ghassan AlRegib

The widespread adoption of deep neural networks in machine learning calls for an objective quantification of esoteric trust. In this paper we propose GradTrust, a classification trust measure for large-scale neural networks at inference. The proposed method utilizes variance of counterfactual gradients, i.e. the required changes in the network parameters if the label were different. We show that GradTrust is superior to existing techniques for detecting misprediction rates on $50000$ images from ImageNet validation dataset. Depending on the network, GradTrust detects images where either the ground truth is incorrect or ambiguous, or the classes are co-occurring. We extend GradTrust to Video Action Recognition on Kinetics-400 dataset. We showcase results on $14$ architectures pretrained on ImageNet and $5$ architectures pretrained on Kinetics-400. We observe the following: (i) simple methodologies like negative log likelihood and margin classifiers outperform state-of-the-art uncertainty and out-of-distribution detection techniques for misprediction rates, and (ii) the proposed GradTrust is in the Top-2 performing methods on $37$ of the considered $38$ experimental modalities. The code is available at: https://github.com/olivesgatech/GradTrust

Mohit Prabhushankar, Ghassan AlRegib

In this paper, we visualize and quantify the predictive uncertainty of gradient-based post hoc visual explanations for neural networks. Predictive uncertainty refers to the variability in the network predictions under perturbations to the input. Visual post hoc explainability techniques highlight features within an image to justify a network's prediction. We theoretically show that existing evaluation strategies of visual explanatory techniques partially reduce the predictive uncertainty of neural networks. This analysis allows us to construct a plug in approach to visualize and quantify the remaining predictive uncertainty of any gradient-based explanatory technique. We show that every image, network, prediction, and explanatory technique has a unique uncertainty. The proposed uncertainty visualization and quantification yields two key observations. Firstly, oftentimes under incorrect predictions, explanatory techniques are uncertain about the same features that they are attributing the predictions to, thereby reducing the trustworthiness of the explanation. Secondly, objective metrics of an explanation's uncertainty, empirically behave similarly to epistemic uncertainty. We support these observations on two datasets, four explanatory techniques, and six neural network architectures. The code is available at https://github.com/olivesgatech/VOICE-Uncertainty.

Ghassan AlRegib, Mohit Prabhushankar

In this article, we present a leap-forward expansion to the study of explainability in neural networks by considering explanations as answers to abstract reasoning-based questions. With $P$ as the prediction from a neural network, these questions are `Why P?', `What if not P?', and `Why P, rather than Q?' for a given contrast prediction $Q$. The answers to these questions are observed correlations, observed counterfactuals, and observed contrastive explanations respectively. Together, these explanations constitute the abductive reasoning scheme. We term the three explanatory schemes as observed explanatory paradigms. The term observed refers to the specific case of post-hoc explainability, when an explanatory technique explains the decision $P$ after a trained neural network has made the decision $P$. The primary advantage of viewing explanations through the lens of abductive reasoning-based questions is that explanations can be used as reasons while making decisions. The post-hoc field of explainability, that previously only justified decisions, becomes active by being involved in the decision making process and providing limited, but relevant and contextual interventions. The contributions of this article are: ($i$) realizing explanations as reasoning paradigms, ($ii$) providing a probabilistic definition of observed explanations and their completeness, ($iii$) creating a taxonomy for evaluation of explanations, and ($iv$) positioning gradient-based complete explanainability's replicability and reproducibility across multiple applications and data modalities, ($v$) code repositories, publicly available at https://github.com/olivesgatech/Explanatory-Paradigms.

Kiran Kokilepersaud, Seulgi Kim, Mohit Prabhushankar et al.

In this paper, we propose an algorithm that can be used on top of a wide variety of self-supervised (SSL) approaches to take advantage of hierarchical structures that emerge during training. SSL approaches typically work through some invariance term to ensure consistency between similar samples and a regularization term to prevent global dimensional collapse. Dimensional collapse refers to data representations spanning a lower-dimensional subspace. Recent work has demonstrated that the representation space of these algorithms gradually reflects a semantic hierarchical structure as training progresses. Data samples of the same hierarchical grouping tend to exhibit greater dimensional collapse locally compared to the dataset as a whole due to sharing features in common with each other. Ideally, SSL algorithms would take advantage of this hierarchical emergence to have an additional regularization term to account for this local dimensional collapse effect. However, the construction of existing SSL algorithms does not account for this property. To address this, we propose an adaptive algorithm that performs a weighted decomposition of the denominator of the InfoNCE loss into two terms: local hierarchical and global collapse regularization respectively. This decomposition is based on an adaptive threshold that gradually lowers to reflect the emerging hierarchical structure of the representation space throughout training. It is based on an analysis of the cosine similarity distribution of samples in a batch. We demonstrate that this hierarchical emergence exploitation (HEX) approach can be integrated across a wide variety of SSL algorithms. Empirically, we show performance improvements of up to 5.6% relative improvement over baseline SSL approaches on classification accuracy on Imagenet with 100 epochs of training.

Kiran Kokilepersaud, Mohit Prabhushankar, Ghassan AlRegib et al.

In this paper, we propose a novel selection strategy for contrastive learning for medical images. On natural images, contrastive learning uses augmentations to select positive and negative pairs for the contrastive loss. However, in the medical domain, arbitrary augmentations have the potential to distort small localized regions that contain the biomarkers we are interested in detecting. A more intuitive approach is to select samples with similar disease severity characteristics, since these samples are more likely to have similar structures related to the progression of a disease. To enable this, we introduce a method that generates disease severity labels for unlabeled OCT scans on the basis of gradient responses from an anomaly detection algorithm. These labels are used to train a supervised contrastive learning setup to improve biomarker classification accuracy by as much as 6% above self-supervised baselines for key indicators of Diabetic Retinopathy.

Johannes Schneider, Mohit Prabhushankar

The learning dynamics of deep neural networks are not well understood. The information bottleneck (IB) theory proclaimed separate fitting and compression phases. But they have since been heavily debated. We comprehensively analyze the learning dynamics by investigating a layer's reconstruction ability of the input and prediction performance based on the evolution of parameters during training. We empirically show the existence of three phases using common datasets and architectures such as ResNet and VGG: (i) near constant reconstruction loss, (ii) decrease, and (iii) increase. We also derive an empirically grounded data model and prove the existence of phases for single-layer networks. Technically, our approach leverages classical complexity analysis. It differs from IB by relying on measuring reconstruction loss rather than information theoretic measures to relate information of intermediate layers and inputs. Our work implies a new best practice for transfer learning: We show empirically that the pre-training of a classifier should stop well before its performance is optimal.

Ghazal Kaviani, Yavuz Yarici, Seulgi Kim et al.

Daily Activity Recordings for Artificial Intelligence (DARai, pronounced "Dahr-ree") is a multimodal, hierarchically annotated dataset constructed to understand human activities in real-world settings. DARai consists of continuous scripted and unscripted recordings of 50 participants in 10 different environments, totaling over 200 hours of data from 20 sensors including multiple camera views, depth and radar sensors, wearable inertial measurement units (IMUs), electromyography (EMG), insole pressure sensors, biomonitor sensors, and gaze tracker. To capture the complexity in human activities, DARai is annotated at three levels of hierarchy: (i) high-level activities (L1) that are independent tasks, (ii) lower-level actions (L2) that are patterns shared between activities, and (iii) fine-grained procedures (L3) that detail the exact execution steps for actions. The dataset annotations and recordings are designed so that 22.7% of L2 actions are shared between L1 activities and 14.2% of L3 procedures are shared between L2 actions. The overlap and unscripted nature of DARai allows counterfactual activities in the dataset. Experiments with various machine learning models showcase the value of DARai in uncovering important challenges in human-centered applications. Specifically, we conduct unimodal and multimodal sensor fusion experiments for recognition, temporal localization, and future action anticipation across all hierarchical annotation levels. To highlight the limitations of individual sensors, we also conduct domain-variant experiments that are enabled by DARai's multi-sensor and counterfactual activity design setup. The code, documentation, and dataset are available at the dedicated DARai website: https://alregib.ece.gatech.edu/software-and-datasets/darai-daily-activity-recordings-for-artificial-intelligence-and-machine-learning/

Chen Zhou, Mohit Prabhushankar, Ghassan AlRegib

Annotators exhibit disagreement during data labeling, which can be termed as annotator label uncertainty. Annotator label uncertainty manifests in variations of labeling quality. Training with a single low-quality annotation per sample induces model reliability degradations. In this work, we first examine the effects of annotator label uncertainty in terms of the model's generalizability and prediction uncertainty. We observe that the model's generalizability and prediction uncertainty degrade with the presence of low-quality noisy labels. Meanwhile, our evaluation of existing uncertainty estimation algorithms indicates their incapability in response to annotator label uncertainty. To mitigate performance degradation, prior methods show that training models with labels collected from multiple independent annotators can enhance generalizability. However, they require massive annotations. Hence, we introduce a novel perceptual quality-based model training framework to objectively generate multiple labels for model training to enhance reliability, while avoiding massive annotations. Specifically, we first select a subset of samples with low perceptual quality scores ranked by statistical regularities of visual signals. We then assign de-aggregated labels to each sample in this subset to obtain a training set with multiple labels. Our experiments and analysis demonstrate that training with the proposed framework alleviates the degradation of generalizability and prediction uncertainty caused by annotator label uncertainty.

Yavuz Yarici, Kiran Kokilepersaud, Mohit Prabhushankar et al.

The high cost of annotating data makes self-supervised approaches, such as contrastive learning methods, appealing for Human Activity Recognition (HAR). Effective contrastive learning relies on selecting informative positive and negative samples. However, HAR sensor signals are subject to significant domain shifts caused by subject variability. These domain shifts hinder model generalization to unseen subjects by embedding subject-specific variations rather than activity-specific features. As a result, human activity recognition models trained with contrastive learning often struggle to generalize to new subjects. We introduce Subject-Invariant Contrastive Learning (SICL), a simple yet effective loss function to improve generalization in human activity recognition. SICL re-weights negative pairs drawn from the same subject to suppress subject-specific cues and emphasize activity-specific information. We evaluate our loss function on three public benchmarks: UTD-MHAD, MMAct, and DARai. We show that SICL improves performance by up to 11% over traditional contrastive learning methods. Additionally, we demonstrate the adaptability of our loss function across various settings, including multiple self-supervised methods, multimodal scenarios, and supervised learning frameworks.

Kiran Kokilepersaud, Mohit Prabhushankar, Ghassan AlRegib

A key factor in effective Self-Supervised learning (SSL) is preventing dimensional collapse, where higher-dimensional representation spaces ($R$) span a lower-dimensional subspace. Therefore, SSL optimization strategies involve guiding a model to produce $R$ with a higher dimensionality ($H(R)$) through objectives that encourage decorrelation of features or sample uniformity in $R$. A higher $H(R)$ indicates that $R$ has greater feature diversity which is useful for generalization to downstream tasks. Alongside dimensionality optimization, SSL algorithms also utilize a projection head that maps $R$ into an embedding space $Z$. Recent work has characterized the projection head as a filter of noisy or irrelevant features from the SSL objective by reducing the mutual information $I(R;Z)$. Therefore, the current literature's view is that a good SSL representation space should have a high $H(R)$ and a low $I(R;Z)$. However, this view of SSL is lacking in terms of an understanding of the underlying training dynamics that influences the relationship between both terms. Our analysis shows that the best performing SSL models do not have the highest $H(R)$ nor the lowest $I(R;Z)$, but effectively arrive at a balance between both. To take advantage of this analysis, we introduce AdaDim, a training strategy that leverages SSL training dynamics by adaptively balancing between increasing $H(R)$ through feature decorrelation and sample uniformity as well as gradual regularization of $I(R;Z)$ as training progresses. We show performance improvements of up to 3% over common SSL baselines despite our method not utilizing expensive techniques such as queues, clustering, predictor networks, or student-teacher architectures.

Jorge Quesada, Chen Zhou, Prithwijit Chowdhury et al.

Machine learning has taken a critical role in seismic interpretation workflows, especially in fault delineation tasks. However, despite the recent proliferation of pretrained models and synthetic datasets, the field still lacks a systematic understanding of the generalizability limits of these models across seismic data representing diverse geologic, acquisition and processing settings. Distributional shifts between data sources, limitations in fine-tuning strategies and labeled data accessibility, and inconsistent evaluation protocols all remain major roadblocks to deploying reliable models in real-world exploration. In this paper, we present the first large-scale benchmarking study explicitly designed to provide guidelines for domain shift strategies in seismic interpretation. Our benchmark spans over 200 combinations of model architectures, datasets and training strategies, across three datasets (synthetic and real) including FaultSeg3D, CRACKS, and Thebe. We systematically assess pretraining, fine-tuning, and joint training under varying domain shifts. Our analysis shows that common fine-tuning practices can lead to catastrophic forgetting, especially when source and target datasets are disjoint, and that larger models such as Segformer are more robust than smaller architectures. We also find that domain adaptation methods outperform fine-tuning when shifts are large, yet underperform when domains are similar. Finally, we complement segmentation metrics with a novel analysis based on fault characteristic descriptors, revealing how models absorb structural biases from training datasets. Overall, we establish a robust experimental baseline that provides insights into tradeoffs in current fault delineation workflows and highlights directions for building more generalizable and interpretable models.

Ryan Benkert, Mohit Prabhushankar, Ghassan AlRegib

The performance of active learning algorithms can be improved in two ways. The often used and intuitive way is by reducing the overall error rate within the test set. The second way is to ensure that correct predictions are not forgotten when the training set is increased in between rounds. The former is measured by the accuracy of the model and the latter is captured in negative flips between rounds. Negative flips are samples that are correctly predicted when trained with the previous/smaller dataset and incorrectly predicted after additional samples are labeled. In this paper, we discuss improving the performance of active learning algorithms both in terms of prediction accuracy and negative flips. The first observation we make in this paper is that negative flips and overall error rates are decoupled and reducing one does not necessarily imply that the other is reduced. Our observation is important as current active learning algorithms do not consider negative flips directly and implicitly assume the opposite. The second observation is that performing targeted active learning on subsets of the unlabeled pool has a significant impact on the behavior of the active learning algorithm and influences both negative flips and prediction accuracy. We then develop ROSE - a plug-in algorithm that utilizes a small labeled validation set to restrict arbitrary active learning acquisition functions to negative flips within the unlabeled pool. We show that integrating a validation set results in a significant performance boost in terms of accuracy, negative flip rate reduction, or both.

Efe Ozturk, Mohit Prabhushankar, Ghassan AlRegib

In this study, we introduce an intelligent Test Time Augmentation (TTA) algorithm designed to enhance the robustness and accuracy of image classification models against viewpoint variations. Unlike traditional TTA methods that indiscriminately apply augmentations, our approach intelligently selects optimal augmentations based on predictive uncertainty metrics. This selection is achieved via a two-stage process: the first stage identifies the optimal augmentation for each class by evaluating uncertainty levels, while the second stage implements an uncertainty threshold to determine when applying TTA would be advantageous. This methodological advancement ensures that augmentations contribute to classification more effectively than a uniform application across the dataset. Experimental validation across several datasets and neural network architectures validates our approach, yielding an average accuracy improvement of 1.73% over methods that use single-view images. This research underscores the potential of adaptive, uncertainty-aware TTA in improving the robustness of image classification in the presence of viewpoint variations, paving the way for further exploration into intelligent augmentation strategies.

Prithwijit Chowdhury, Mohit Prabhushankar, Ghassan AlRegib et al.

Explainable AI (XAI) has revolutionized the field of deep learning by empowering users to have more trust in neural network models. The field of XAI allows users to probe the inner workings of these algorithms to elucidate their decision-making processes. The rise in popularity of XAI has led to the advent of different strategies to produce explanations, all of which only occasionally agree. Thus several objective evaluation metrics have been devised to decide which of these modules give the best explanation for specific scenarios. The goal of the paper is twofold: (i) we employ the notions of necessity and sufficiency from causal literature to come up with a novel explanatory technique called SHifted Adversaries using Pixel Elimination(SHAPE) which satisfies all the theoretical and mathematical criteria of being a valid explanation, (ii) we show that SHAPE is, infact, an adversarial explanation that fools causal metrics that are employed to measure the robustness and reliability of popular importance based visual XAI methods. Our analysis shows that SHAPE outperforms popular explanatory techniques like GradCAM and GradCAM++ in these tests and is comparable to RISE, raising questions about the sanity of these metrics and the need for human involvement for an overall better evaluation.

Yavuz Yarici, Kiran Kokilepersaud, Mohit Prabhushankar et al.

Self-supervised models create representation spaces that lack clear semantic meaning. This interpretability problem of representations makes traditional explainability methods ineffective in this context. In this paper, we introduce a novel method to analyze representation spaces using three key perceptual components: color, shape, and texture. We employ selective masking of these components to observe changes in representations, resulting in distinct importance maps for each. In scenarios, where labels are absent, these importance maps provide more intuitive explanations as they are integral to the human visual system. Our approach enhances the interpretability of the representation space, offering explanations that resonate with human visual perception. We analyze how different training objectives create distinct representation spaces using perceptual components. Additionally, we examine the representation of images across diverse image domains, providing insights into the role of these components in different contexts.

Kiran Kokilepersaud, Yavuz Yarici, Mohit Prabhushankar et al.

In this work, we propose a novel supervised contrastive loss that enables the integration of taxonomic hierarchy information during the representation learning process. A supervised contrastive loss operates by enforcing that images with the same class label (positive samples) project closer to each other than images with differing class labels (negative samples). The advantage of this approach is that it directly penalizes the structure of the representation space itself. This enables greater flexibility with respect to encoding semantic concepts. However, the standard supervised contrastive loss only enforces semantic structure based on the downstream task (i.e. the class label). In reality, the class label is only one level of a \emph{hierarchy of different semantic relationships known as a taxonomy}. For example, the class label is oftentimes the species of an animal, but between different classes there are higher order relationships such as all animals with wings being ``birds". We show that by explicitly accounting for these relationships with a weighting penalty in the contrastive loss we can out-perform the supervised contrastive loss. Additionally, we demonstrate the adaptability of the notion of a taxonomy by integrating our loss into medical and noise-based settings that show performance improvements by as much as 7%.

Ryan Benkert, Mohit Prabhushankar, Ghassan AlRegib

This paper presents a discussion on data selection for deep learning in the field of seismic interpretation. In order to achieve a robust generalization to the target volume, it is crucial to identify the specific samples are the most informative to the training process. The selection of the training set from a target volume is a critical factor in determining the effectiveness of the deep learning algorithm for interpreting seismic volumes. This paper proposes the inclusion of interpretation disagreement as a valuable and intuitive factor in the process of selecting training sets. The development of a novel data selection framework is inspired by established practices in seismic interpretation. The framework we have developed utilizes representation shifts to effectively model interpretation disagreement within neural networks. Additionally, it incorporates the disagreement measure to enhance attention towards geologically interesting regions throughout the data selection workflow. By combining this approach with active learning, a well-known machine learning paradigm for data selection, we arrive at a comprehensive and innovative framework for training set selection in seismic interpretation. In addition, we offer a specific implementation of our proposed framework, which we have named ATLAS. This implementation serves as a means for data selection. In this study, we present the results of our comprehensive experiments, which clearly indicate that ATLAS consistently surpasses traditional active learning frameworks in the field of seismic interpretation. Our findings reveal that ATLAS achieves improvements of up to 12% in mean intersection-over-union.

Kiran Kokilepersaud, Stephanie Trejo Corona, Mohit Prabhushankar et al.

This paper presents a novel positive and negative set selection strategy for contrastive learning of medical images based on labels that can be extracted from clinical data. In the medical field, there exists a variety of labels for data that serve different purposes at different stages of a diagnostic and treatment process. Clinical labels and biomarker labels are two examples. In general, clinical labels are easier to obtain in larger quantities because they are regularly collected during routine clinical care, while biomarker labels require expert analysis and interpretation to obtain. Within the field of ophthalmology, previous work has shown that clinical values exhibit correlations with biomarker structures that manifest within optical coherence tomography (OCT) scans. We exploit this relationship by using the clinical data as pseudo-labels for our data without biomarker labels in order to choose positive and negative instances for training a backbone network with a supervised contrastive loss. In this way, a backbone network learns a representation space that aligns with the clinical data distribution available. Afterwards, we fine-tune the network trained in this manner with the smaller amount of biomarker labeled data with a cross-entropy loss in order to classify these key indicators of disease directly from OCT scans. We also expand on this concept by proposing a method that uses a linear combination of clinical contrastive losses. We benchmark our methods against state of the art self-supervised methods in a novel setting with biomarkers of varying granularity. We show performance improvements by as much as 5\% in total biomarker detection AUROC.

Mohit Prabhushankar, Ghassan AlRegib

Neural networks represent data as projections on trained weights in a high dimensional manifold. The trained weights act as a knowledge base consisting of causal class dependencies. Inference built on features that identify these dependencies is termed as feed-forward inference. Such inference mechanisms are justified based on classical cause-to-effect inductive reasoning models. Inductive reasoning based feed-forward inference is widely used due to its mathematical simplicity and operational ease. Nevertheless, feed-forward models do not generalize well to untrained situations. To alleviate this generalization challenge, we propose using an effect-to-cause inference model that reasons abductively. Here, the features represent the change from existing weight dependencies given a certain effect. We term this change as contrast and the ensuing reasoning mechanism as contrastive reasoning. In this paper, we formalize the structure of contrastive reasoning and propose a methodology to extract a neural network's notion of contrast. We demonstrate the value of contrastive reasoning in two stages of a neural network's reasoning pipeline : in inferring and visually explaining decisions for the application of object recognition. We illustrate the value of contrastively recognizing images under distortions by reporting an improvement of 3.47%, 2.56%, and 5.48% in average accuracy under the proposed contrastive framework on CIFAR-10C, noisy STL-10, and VisDA datasets respectively.

Mohit Prabhushankar, Ghassan AlRegib

Neural networks trained to classify images do so by identifying features that allow them to distinguish between classes. These sets of features are either causal or context dependent. Grad-CAM is a popular method of visualizing both sets of features. In this paper, we formalize this feature divide and provide a methodology to extract causal features from Grad-CAM. We do so by defining context features as those features that allow contrast between predicted class and any contrast class. We then apply a set theoretic approach to separate causal from contrast features for COVID-19 CT scans. We show that on average, the image regions with the proposed causal features require 15% less bits when encoded using Huffman encoding, compared to Grad-CAM, for an average increase of 3% classification accuracy, over Grad-CAM. Moreover, we validate the transfer-ability of causal features between networks and comment on the non-human interpretable causal nature of current networks.

Gukyeong Kwon, Mohit Prabhushankar, Dogancan Temel et al.

In this paper, we propose a model-based characterization of neural networks to detect novel input types and conditions. Novelty detection is crucial to identify abnormal inputs that can significantly degrade the performance of machine learning algorithms. Majority of existing studies have focused on activation-based representations to detect abnormal inputs, which limits the characterization of abnormality from a data perspective. However, a model perspective can also be informative in terms of the novelties and abnormalities. To articulate the significance of the model perspective in novelty detection, we utilize backpropagated gradients. We conduct a comprehensive analysis to compare the representation capability of gradients with that of activation and show that the gradients outperform the activation in novel class and condition detection. We validate our approach using four image recognition datasets including MNIST, Fashion-MNIST, CIFAR-10, and CURE-TSR. We achieve a significant improvement on all four datasets with an average AUROC of 0.953, 0.918, 0.582, and 0.746, respectively.

Yutong Sun, Mohit Prabhushankar, Ghassan AlRegib

In this paper, we show that existing recognition and localization deep architectures, that have not been exposed to eye tracking data or any saliency datasets, are capable of predicting the human visual saliency. We term this as implicit saliency in deep neural networks. We calculate this implicit saliency using expectancy-mismatch hypothesis in an unsupervised fashion. Our experiments show that extracting saliency in this fashion provides comparable performance when measured against the state-of-art supervised algorithms. Additionally, the robustness outperforms those algorithms when we add large noise to the input images. Also, we show that semantic features contribute more than low-level features for human visual saliency detection.

Mohit Prabhushankar, Gukyeong Kwon, Dogancan Temel et al.

Visual explanations are logical arguments based on visual features that justify the predictions made by neural networks. Current modes of visual explanations answer questions of the form $`Why \text{ } P?'$. These $Why$ questions operate under broad contexts thereby providing answers that are irrelevant in some cases. We propose to constrain these $Why$ questions based on some context $Q$ so that our explanations answer contrastive questions of the form $`Why \text{ } P, \text{} rather \text{ } than \text{ } Q?'$. In this paper, we formalize the structure of contrastive visual explanations for neural networks. We define contrast based on neural networks and propose a methodology to extract defined contrasts. We then use the extracted contrasts as a plug-in on top of existing $`Why \text{ } P?'$ techniques, specifically Grad-CAM. We demonstrate their value in analyzing both networks and data in applications of large-scale recognition, fine-grained recognition, subsurface seismic analysis, and image quality assessment.

Gukyeong Kwon, Mohit Prabhushankar, Dogancan Temel et al.

Learning representations that clearly distinguish between normal and abnormal data is key to the success of anomaly detection. Most of existing anomaly detection algorithms use activation representations from forward propagation while not exploiting gradients from backpropagation to characterize data. Gradients capture model updates required to represent data. Anomalies require more drastic model updates to fully represent them compared to normal data. Hence, we propose the utilization of backpropagated gradients as representations to characterize model behavior on anomalies and, consequently, detect such anomalies. We show that the proposed method using gradient-based representations achieves state-of-the-art anomaly detection performance in benchmark image recognition datasets. Also, we highlight the computational efficiency and the simplicity of the proposed method in comparison with other state-of-the-art methods relying on adversarial networks or autoregressive models, which require at least 27 times more model parameters than the proposed method.

Gukyeong Kwon, Mohit Prabhushankar, Dogancan Temel et al.

In this paper, we utilize weight gradients from backpropagation to characterize the representation space learned by deep learning algorithms. We demonstrate the utility of such gradients in applications including perceptual image quality assessment and out-of-distribution classification. The applications are chosen to validate the effectiveness of gradients as features when the test image distribution is distorted from the train image distribution. In both applications, the proposed gradient based features outperform activation features. In image quality assessment, the proposed approach is compared with other state of the art approaches and is generally the top performing method on TID 2013 and MULTI-LIVE databases in terms of accuracy, consistency, linearity, and monotonic behavior. Finally, we analyze the effect of regularization on gradients using CURE-TSR dataset for out-of-distribution classification.

Mohit Prabhushankar, Gukyeong Kwon, Dogancan Temel et al.

In this paper, we generate and control semantically interpretable filters that are directly learned from natural images in an unsupervised fashion. Each semantic filter learns a visually interpretable local structure in conjunction with other filters. The significance of learning these interpretable filter sets is demonstrated on two contrasting applications. The first application is image recognition under progressive decolorization, in which recognition algorithms should be color-insensitive to achieve a robust performance. The second application is image quality assessment where objective methods should be sensitive to color degradations. In the proposed work, the sensitivity and lack thereof are controlled by weighing the semantic filters based on the local structures they represent. To validate the proposed approach, we utilize the CURE-TSR dataset for image recognition and the TID 2013 dataset for image quality assessment. We show that the proposed semantic filter set achieves state-of-the-art performances in both datasets while maintaining its robustness across progressive distortions.

Mohit Prabhushankar, Dogancan Temel, Ghassan AlRegib

In this paper, we train independent linear decoder models to estimate the perceived quality of images. More specifically, we calculate the responses of individual non-overlapping image patches to each of the decoders and scale these responses based on the sharpness characteristics of filter set. We use multiple linear decoders to capture different abstraction levels of the image patches. Training each model is carried out on 100,000 image patches from the ImageNet database in an unsupervised fashion. Color space selection and ZCA Whitening are performed over these patches to enhance the descriptiveness of the data. The proposed quality estimator is tested on the LIVE and the TID 2013 image quality assessment databases. Performance of the proposed method is compared against eleven other state of the art methods in terms of accuracy, consistency, linearity, and monotonic behavior. Based on experimental results, the proposed method is generally among the top performing quality estimators in all categories.

Mohit Prabhushankar, Dogancan Temel, Ghassan AlRegib

In this paper, we introduce an adaptive unsupervised learning framework, which utilizes natural images to train filter sets. The applicability of these filter sets is demonstrated by evaluating their performance in two contrasting applications - image quality assessment and texture retrieval. While assessing image quality, the filters need to capture perceptual differences based on dissimilarities between a reference image and its distorted version. In texture retrieval, the filters need to assess similarity between texture images to retrieve closest matching textures. Based on experiments, we show that the filter responses span a set in which a monotonicity-based metric can measure both the perceptual dissimilarity of natural images and the similarity of texture images. In addition, we corrupt the images in the test set and demonstrate that the proposed method leads to robust and reliable retrieval performance compared to existing methods.

Dogancan Temel, Gukyeong Kwon, Mohit Prabhushankar et al.

In this paper, we investigate the robustness of traffic sign recognition algorithms under challenging conditions. Existing datasets are limited in terms of their size and challenging condition coverage, which motivated us to generate the Challenging Unreal and Real Environments for Traffic Sign Recognition (CURE-TSR) dataset. It includes more than two million traffic sign images that are based on real-world and simulator data. We benchmark the performance of existing solutions in real-world scenarios and analyze the performance variation with respect to challenging conditions. We show that challenging conditions can decrease the performance of baseline methods significantly, especially if these challenging conditions result in loss or misplacement of spatial information. We also investigate the effect of data augmentation and show that utilization of simulator data along with real-world data enhance the average recognition performance in real-world scenarios. The dataset is publicly available at https://ghassanalregib.com/cure-tsr/.