Welcome

I am a PhD student in the Computer Vision and Machine Learning group at the Max Planck Institute for Informatics in the Saarland Informatics Campus, supervised by Prof. Bernt Schiele. I am also an associated member of the RTG Neuroexplicit Models of Language, Vision, and Action.

I am interested in understanding the inner workings and mechanisms of deep models—recently, specifically multimodal vision-language models—and using this understanding to build trustworthy, performant, and safe systems.

I completed the preparatory phase of my PhD at the Max Planck Graduate Center for Computer and Information Science, during which I worked on improving adversarial robustness of image classifiers against adversarial patch attacks. I received my B.Tech. in Computer Science and Engineering with the President of India Gold Medal from the Indian Institute of Technology Hyderabad in August 2019. At IIT Hyderabad, I worked on projects for improving crowd-sourced vote aggregation supervised by Prof. Vineeth N Balasubramanian; and for developing state-of-the-art incomplete MaxSAT solvers with Prof. Saurabh Joshi and Prof. Ruben Martins. Within this period, I also worked as an intern in the Machine Intelligence Laboratory at The University of Tokyo under Prof. Tatsuya Harada on visual relationship detection, and at Bosch on classification of fundus images using deep learning.

News

Publications

Google Scholar DBLP

Foundational vision models have become the de facto standard for many vision tasks due to their strong performance. However, they are notoriously opaque and remain hard to interpret. We present ALOE (ALign Once to Explain), a one-time, label-free feature alignment approach that efficiently converts foundational vision models into inherently interpretable B-cos variants. Once aligned, the B-cos backbone is used as a drop-in replacement across several downstream tasks—amortizing the cost of interpretability. ALOE is robust across pre-training paradigms (supervised, self-supervised, vision–language) and is 100–1000x more data-efficient than training from scratch. On classification, it strongly outperforms vanilla B-cosification (e.g., +9.2 p.p. top-1 on ImageNet for supervised ViT-B/16), retains strong linear probing, k-NN, and zero-shot transfer competitive with foundational backbones (DINOv3, SigLIP2) across diverse downstream datasets. It also preserves spatially structured features useful for dense prediction, while yielding well-localized and highly human-interpretable explanations by design.

Align Once to Explain: Feature Alignment for Scalable B-cosification of Foundational Vision Transformers

Raphael Maser*, Siddhartha Gairola*, Sukrut Rao, Bernt Schiele
CVPR 2026
Paper Poster Video Slides
ALOE is a method to transform large-scale ViT-based foundation models such as DINOv3 and SigLIP2 into inherently interpretable B-cos variants at a fraction of the cost of training from scratch, bringing interpretability while retaining strong performance across a range of downstream tasks and datasets.
Language-aligned vision foundation models perform strongly across diverse downstream tasks. Yet, their learned representations remain opaque, making interpreting their decision-making difficult. Recent work decompose these representations into human-interpretable concepts, but provide poor spatial grounding and are limited to image classification tasks. In this work, we propose CFM, a language-aligned concept foundation model for vision that provides fine-grained concepts, which are human-interpretable and spatially grounded in the input image. When paired with a foundation model with strong semantic representations, we get explanations for any of its downstream tasks. Examining local co-occurrence dependencies of concepts allows us to define concept relationships through which we improve concept naming and obtain richer explanations. On benchmark data, we show that CFM provides performance on classification, segmentation, and captioning that is competitive with opaque foundation models while providing fine-grained, high quality concept-based explanations. Code at https://github.com/kawi19/CFM.

CFM: Language-aligned Concept Foundation Model for Vision

Kai Wittenmayer, Sukrut Rao, Amin Parchami-Araghi, Bernt Schiele, Jonas Fischer
arXiv 2026
DOI arXiv Code
CFM is a language-aligned concept foundation model that extracts spatially localized, visually grounded, and human-interpretable concepts at various granularities from images, organizes them into hierarchies, automatically assigns names to them, enabling concept-based explanations for any downstream task that the foundation model can perform, such as classification, open vocabulary segmentation, and captioning.
Deep networks have shown remarkable performance across a wide range of tasks, yet getting a global concept-level understanding of how they function remains a key challenge. Many post-hoc concept-based approaches have been introduced to understand their workings, yet they are not always faithful to the model. Further, they make restrictive assumptions on the concepts a model learns, such as class-specificity, small spatial extent, or alignment to human expectations. In this work, we put emphasis on the faithfulness of such concept-based explanations and propose a new model with model-inherent mechanistic concept-explanations. Our concepts are shared across classes and, from any layer, their contribution to the logit and their input-visualization can be faithfully traced. We also leverage foundation models to propose a new concept-consistency metric, C -Score, that can be used to evaluate concept-based methods. We show that, compared to prior work, our concepts are quantitatively more consistent and users find our concepts to be more interpretable, all while retaining competitive ImageNet performance.

FaCT: Faithful Concept Traces for Explaining Neural Network Decisions

Amin Parchami-Araghi, Sukrut Rao, Jonas Fischer, Bernt Schiele
NeurIPS 2025
DOI arXiv Paper Code Poster Video Slides
FaCT combines concept-discovery with model-inherent attributions to construct a model that provides faithful concept traces for explaining its decisions, i.e., contributions of pixels to concepts and concepts to the final decision can be faithfully traced. We also propose a novel concept-consistency metric, C2-Score, and show that FaCT yields more consistent and interpretable concepts while retaining competitive performance.
Post-hoc explanation methods for black-box models often struggle with faithfulness and human interpretability due to the lack of explainability in current neural architectures. Meanwhile, B-cos networks have been introduced to improve model explainability by proposing an architecture that removes bias terms and promotes input-weight alignment. Although B-cos networks have shown success in building explainable systems, their application has so far been limited to computer vision models and their associated training pipelines. In this work, we introduce B-cos LMs, i.e., B-cos Language Models (LMs) empowered for natural language processing (NLP) tasks. Our approach directly transforms pre-trained language models into B-cos LMs by combining B-cos conversion and task fine-tuning, improving efficiency compared to previous methods. Automatic and human evaluation results demonstrate that B-cos LMs produce more faithful and human interpretable explanations than post-hoc methods, while maintaining task performance comparable to conventional fine-tuning. Our in-depth analysis explores how B-cos LMs differ from conventionally fine-tuned models in their learning processes and explanation patterns. Finally, we present a first exploration of transforming decoder-only models to B-cos LMs for generation tasks. Our code is available at https://github.com/Ewanwong/bcos_lm.

B-cos LM: Efficiently Transforming Pre-trained Language Models for Improved Explainability

Yifan Wang, Sukrut Rao, Ji-Ung Lee, Mayank Jobanputra, Vera Demberg
TMLR 2025
DOI arXiv Paper Code
B-cos LMs extend B-cos networks to language models, providing more faithful and human interpretable explanations than post-hoc methods while maintaining comparable task performance.
B-cos Networks have been shown to be effective for obtaining highly human interpretable explanations of model decisions by architecturally enforcing stronger alignment between inputs and weight. B-cos variants of convolutional networks (CNNs) and vision transformers (ViTs), which primarily replace linear layers with B-cos transformations, perform competitively to their respective standard variants while also yielding explanations that are faithful by design. However, it has so far been necessary to train these models from scratch, which is increasingly infeasible in the era of large, pre-trained foundation models. In this work, inspired by the architectural similarities in standard DNNs and B-cos networks, we propose ‘B-cosification’, a novel approach to transform existing pre-trained models to become inherently interpretable. We perform a thorough study of design choices to perform this conversion, both for convolutional neural networks and vision transformers. We find that B-cosification can yield models that are on par with B-cos models trained from scratch in terms of interpretability, while often outperforming them in terms of classification performance at a fraction of the training cost. Subsequently, we apply B-cosification to a pretrained CLIP model, and show that, even with limited data and compute cost, we obtain a B-cosified version that is highly interpretable and competitive on zero shot performance across a variety of datasets. We release our code and pre-trained model weights at https://github.com/shrebox/B-cosification.

B-cosification: Transforming Deep Neural Networks to be Inherently Interpretable

Shreyash Arya*, Sukrut Rao*, Moritz Böhle*, Bernt Schiele
NeurIPS 2024
DOI arXiv Paper Code Poster Video Slides
B-cosification is a method to transform existing pre-trained models to inherently interpretable B-cos variants at a fraction of the cost of training from scratch, yielding models that are interpretable while often outperforming them in terms of classification performance. We also apply B-cosification to CLIP and show that the B-cosified version remains competitive on performance while being interpretable.
Concept Bottleneck Models (CBMs) have recently been proposed to address the ‘black-box’ problem of deep neural networks, by first mapping images to a human-understandable concept space and then linearly combining concepts for classification. Such models typically require first coming up with a set of concepts relevant to the task and then aligning the representations of a feature extractor to map to these concepts. However, even with powerful foundational feature extractors like CLIP, there are no guarantees that the specified concepts are detectable. In this work, we leverage recent advances in mechanistic interpretability and propose a novel CBM approach — called Discover-then-Name-CBM (DN-CBM) — that inverts the typical paradigm: instead of pre-selecting concepts based on the downstream classification task, we use sparse autoencoders to first discover concepts learnt by the model, and then name them and train linear probes for classification. Our concept extraction strategy is efficient, since it is agnostic to the downstream task, and uses concepts already known to the model. We perform a comprehensive evaluation across multiple datasets and CLIP architectures and show that our method yields semantically meaningful concepts, assigns appropriate names to them that make them easy to interpret, and yields performant and interpretable CBMs. Code available at https://github.com/neuroexplicit-saar/discover-then-name.

Discover-then-Name: Task-Agnostic Concept Bottlenecks via Automated Concept Discovery

Sukrut Rao*, Sweta Mahajan*, Moritz Böhle, Bernt Schiele
ECCV 2024
DOI arXiv Paper Code Poster Video Slides
Discover-then-Name is an efficient task-agnostic approach to build concept bottleneck models (CBMs) by first discovering concepts learnt by the model using sparse autoencoders and then naming them automatically, yielding semantically meaningful concepts with appropriate names that help construct performant and interpretable CBMs.
Knowledge Distillation (KD) has proven effective for compressing large teacher models into smaller student models. While it is well known that student models can achieve similar accuracies as the teachers, it has also been shown that they nonetheless often do not learn the same function. It is, however, often highly desirable that the student's and teacher's functions share similar properties such as basing the prediction on the same input features, as this ensures that students learn the 'right features' from the teachers. In this work, we explore whether this can be achieved by not only optimizing the classic KD loss but also the similarity of the explanations generated by the teacher and the student. Despite the idea being simple and intuitive, we find that our proposed 'explanation-enhanced' KD (e2KD) (1) consistently provides large gains in terms of accuracy and student-teacher agreement, (2) ensures that the student learns from the teacher to be right for the right reasons and to give similar explanations, and (3) is robust with respect to the model architectures, the amount of training data, and even works with 'approximate', pre-computed explanations.

Good Teachers Explain: Explanation-Enhanced Knowledge Distillation

Amin Parchami-Araghi*, Moritz Böhle*, Sukrut Rao*, Bernt Schiele
ECCV 2024
DOI arXiv Paper Code Poster Video Slides
Explanation-enhanced knowledge distillation (e2KD) is a method to faithfully distill teacher models into students by additionally optimizing the similarity of teacher and student explanations. We show that e2KD consistently improves accuracy and student-teacher agreement, ensures that students learn from teachers to be right for the right reasons, and is robust across architectures, data amounts, and works even with pre-computed explanations.
Deep neural networks are very successful on many vision tasks, but hard to interpret due to their black box nature. To overcome this, various post-hoc attribution methods have been proposed to identify image regions most influential to the models' decisions. Evaluating such methods is challenging since no ground truth attributions exist. We thus propose three novel evaluation schemes to more reliably measure the faithfulness of those methods, to make comparisons between them more fair, and to make visual inspection more systematic. To address faithfulness, we propose a novel evaluation setting (DiFull) in which we carefully control which parts of the input can influence the output in order to distinguish possible from impossible attributions. To address fairness, we note that different methods are applied at different layers, which skews any comparison, and so evaluate all methods on the same layers (ML-Att) and discuss how this impacts their performance on quantitative metrics. For more systematic visualizations, we propose a scheme (AggAtt) to qualitatively evaluate the methods on complete datasets. We use these evaluation schemes to study strengths and shortcomings of some widely used attribution methods over a wide range of models. Finally, we propose a post-processing smoothing step that significantly improves the performance of some attribution methods, and discuss its applicability.

Better Understanding Differences in Attribution Methods via Systematic Evaluations

Sukrut Rao, Moritz Böhle, Bernt Schiele
IEEE TPAMI 2024
Project DOI arXiv Paper Code Conference Version
We propose three novel evaluation schemes to better understand the faithfulness and differences between attribution methods, and use them to study strengths and shortcomings of some widely used attribution methods. We extend our work on attribution evaluation to more attribution methods, models, and perform additional analyses.
Despite being highly performant, deep neural networks might base their decisions on features that spuriously correlate with the provided labels, thus hurting generalization. To mitigate this, ‘model guidance’ has recently gained popularity, i.e. the idea of regularizing the models’ explanations to ensure that they are “right for the right reasons”. While various techniques to achieve such model guidance have been proposed, experimental validation of these approaches has thus far been limited to relatively simple and / or synthetic datasets. To better understand the effectiveness of the various design choices that have been explored in the context of model guidance, in this work we conduct an in-depth evaluation across various loss functions, attribution methods, models, and ‘guidance depths’ on the PASCAL VOC 2007 and MS COCO 2014 datasets. As annotation costs for model guidance can limit its applicability, we also place a particular focus on efficiency. Specifically, we guide the models via bounding box annotations, which are much cheaper to obtain than the commonly used segmentation masks, and evaluate the robustness of model guidance under limited (e.g. with only 1% of annotated images) or overly coarse annotations. Further, we propose using the EPG score as an additional evaluation metric and loss function (‘Energy loss’). We show that optimizing for the Energy loss leads to models that exhibit a distinct focus on object-specific features, despite only using bounding box annotations that also include background regions. Lastly, we show that such model guidance can improve generalization under distribution shifts. Code available at: https://github.com/sukrutrao/Model-Guidance

Studying How to Efficiently and Effectively Guide Models with Explanations

Sukrut Rao*, Moritz Böhle*, Amin Parchami-Araghi, Bernt Schiele
ICCV 2023
Project DOI arXiv Paper Code Poster Video
We perform an in-depth study on model guidance with explanations by evaluating various design choices, and also explore ways to improve efficiency. We show that guidance is effective even with limited, coarse, and noisy annotations, using the energy loss with model-inherent B-cos explanations works the best, and that guidance can help improve generalization under distribution shifts.
Deep neural networks are very successful on many vision tasks, but hard to interpret due to their black box nature. To overcome this, various post-hoc attribution methods have been proposed to identify image regions most influential to the models' decisions. Evaluating such methods is challenging since no ground truth attributions exist. We thus propose three novel evaluation schemes to more reliably measure the faithfulness of those methods, to make comparisons between them more fair, and to make visual inspection more systematic. To address faithfulness, we propose a novel evaluation setting (DiFull) in which we carefully control which parts of the input can influence the output in order to distinguish possible from impossible attributions. To address fairness, we note that different methods are applied at different layers, which skews any comparison, and so evaluate all methods on the same layers (ML-Att) and discuss how this impacts their performance on quantitative metrics. For more systematic visualizations, we propose a scheme (AggAtt) to qualitatively evaluate the methods on complete datasets. We use these evaluation schemes to study strengths and shortcomings of some widely used attribution methods. Finally, we propose a post-processing smoothing step that significantly improves the performance of some attribution methods, and discuss its applicability.

Towards Better Understanding Attribution Methods

Sukrut Rao, Moritz Böhle, Bernt Schiele
CVPR 2022
Project DOI arXiv Paper Code Poster Video Journal Version
We propose three novel evaluation schemes to better understand the faithfulness and differences between attribution methods, and use them to study strengths and shortcomings of some widely used attribution methods.
Deep neural networks have been shown to be susceptible to adversarial examples -- small, imperceptible changes constructed to cause mis-classification in otherwise highly accurate image classifiers. As a practical alternative, recent work proposed so-called adversarial patches: clearly visible, but adversarially crafted rectangular patches in images. These patches can easily be printed and applied in the physical world. While defenses against imperceptible adversarial examples have been studied extensively, robustness against adversarial patches is poorly understood. In this work, we first devise a practical approach to obtain adversarial patches while actively optimizing their location within the image. Then, we apply adversarial training on these location-optimized adversarial patches and demonstrate significantly improved robustness on CIFAR10 and GTSRB. Additionally, in contrast to adversarial training on imperceptible adversarial examples, our adversarial patch training does not reduce accuracy.

Adversarial Training against Location-Optimized Adversarial Patches

Sukrut Rao, David Stutz, Bernt Schiele
ECCV-W 2020
Project DOI arXiv Paper Code Slides Video
We show that location-optimization significantly strengthens adversarial patch attacks, and then show that adversarial training on these stronger attacks significantly improves robustness without reducing accuracy
Incomplete MaxSAT solving aims to quickly find a solution that attempts to minimize the sum of the weights of unsatisfied soft clauses without providing any optimality guarantees. In this paper, we propose two approximation strategies for improving incomplete weighted MaxSAT solving. In one of the strategies, we cluster the weights and approximate them with a representative weight. In another strategy, we break up the problem of minimizing the sum of weights of unsatisfiable clauses into multiple minimization subproblems. We have implemented these strategies in a tool Open-WBO-Inc. Using the subproblem minimization strategy, Open-WBO-Inc placed first and second in the weighted incomplete tracks in the MaxSAT Evaluation 2018 whereas the strategy based on weight approximation was placed fourth. We compare these strategies with the best incomplete MaxSAT solvers on benchmarks taken from MaxSAT Evaluation 2017 and MaxSAT Evaluation 2018 and show that the strategies proposed are competitive with the best of the solvers.

Open-WBO-Inc: Approximation Strategies for Incomplete Weighted MaxSAT

Saurabh Joshi, Prateek Kumar, Sukrut Rao, Ruben Martins
JSAT 2019
Project DOI Paper Code Conference Version
We propose two approximation strategies for improving incomplete MaxSAT solving that provide state-of-the-art results on the MaxSAT Evaluation benchmarks.
Incomplete MaxSAT solving aims to quickly find a solution that attempts to minimize the sum of the weights of the unsatisfied soft clauses without providing any optimality guarantees. In this paper, we propose two approximation strategies for improving incomplete MaxSAT solving. In one of the strategies, we cluster the weights and approximate them with a representative weight. In another strategy, we break up the problem of minimizing the sum of weights of unsatisfiable clauses into multiple minimization subproblems. Experimental results show that approximation strategies can be used to find better solutions than the best incomplete solvers in the MaxSAT Evaluation 2017.

Approximation Strategies for Incomplete MaxSAT

Saurabh Joshi, Prateek Kumar, Ruben Martins, Sukrut Rao
CP 2018
Project DOI arXiv Paper Code Slides Journal Version
We propose two approximation strategies for improving incomplete MaxSAT solving that provide state-of-the-art results on the MaxSAT Evaluation benchmarks.
Many real world problems can now be effectively solved using supervised machine learning. A major roadblock is often the lack of an adequate quantity of labeled data for training. A possible solution is to assign the task of labeling data to a crowd, and then infer the true label using aggregation methods. A well-known approach for aggregation is the Dawid-Skene (DS) algorithm, which is based on the principle of Expectation-Maximization (EM). We propose a new simple, yet effective, EM-based algorithm, which can be interpreted as a 'hard' version of DS, that allows much faster convergence while maintaining similar accuracy in aggregation. We show the use of this algorithm as a quick and effective technique for online, real-time sentiment annotation. Our experiments on standard datasets show a significant speedup in time taken for aggregation - upto ~8x over Dawid-Skene and ~6x over other fast EM methods, at competitive accuracy performance.

Fast Dawid-Skene: A Fast Vote Aggregation Scheme for Sentiment Classification

Vaibhav B Sinha, Sukrut Rao, Vineeth N Balasubramanian
KDD WISDOM 2018
Project DOI arXiv Paper Code Slides Web Page
Fast Dawid-Skene is a fast EM-based algorithm for vote aggregation that can be used for real-time sentiment annotation, providing significant speedup over Dawid-Skene and other fast EM methods while maintaining competitive accuracy performance.

Talks

May 2026 Multimodal AI Lab, TU Darmstadt Understanding and Improving Multimodal Models through the Lens of Interpretability Darmstadt, Germany
May 2026 Computer Vision and Multimodal Learning Un-Workshop, Tübingen AI Center Bringing Interpretability to Vision-Language Foundation Models Tübingen, Germany
Oct 2025 Guide Labs Towards Inherent Interpretability for Foundation Models San Francisco, USA Virtual
Jul 2025 IIIT Summer School on AI Explainable Computer Vision: Feature Attributions, Concepts, and Beyond Hyderabad, India Virtual
Nov 2024 Symposium on Explainable Artificial Intelligence Beyond Simple Attributions Keeping the Model in the Loop: Attribution Methods and Beyond Mainz, Germany

Academic Activities

Organizer

Reviewer

Conferences

  • CVPR 2023–2026 (2x Outstanding Reviewer)
  • NeurIPS 2023, 2025–2026 (1x Top Reviewer)
  • ECCV 2024–2026
  • ICML 2024–2026
  • ICLR 2024
  • ICCV 2023–2025

Journals

  • IEEE TPAMI 2024

Workshops

  • NeurIPS XAIA 2023

Miscellaneous

  • Student Volunteer, ICCV 2023

Teaching and Supervision

Master Thesis

Bachelor Thesis

  • Moussa Herrmann, Jul 2025 – Oct 2025
    Refining Concept Bottlenecks with Symbolic Relations
    Co-supervised with Jonas Fischer

Research Immersion / Student Research

Graduate Teaching Assistantships (Saarland University)

  • Explainable Machine Learning Seminar, Winter 2025–26
  • High-Level Computer Vision, Summer 2024
  • High-Level Computer Vision, Summer 2023

Undergraduate Teaching Assistantships (IIT Hyderabad)

  • CS6510: Applied Machine Learning, Spring 2019
  • CS6230: Optimization Methods in Machine Learning, Fall 2018
  • CS3423: Compilers-II, Fall 2018
  • CS2433: Principles of Programming Languages-II, Spring 2018
  • CS2400: Principles of Programming Languages-I, Fall 2017

Awards

Academic Awards

  • Outstanding Reviewer (∼ 5%) at CVPR 2026.
  • Top Reviewer (∼ 10%) at NeurIPS 2023.
  • Outstanding Reviewer (∼ 3.3%) at CVPR 2023.
  • President of India Gold Medal 2019 from IIT Hyderabad.
  • Institute Silver Medal 2019 from IIT Hyderabad.
  • Appreciation in Research Award 2019 from IIT Hyderabad.
  • Honda Young Engineer and Scientist Award 2018 from Honda Foundation, Japan.
  • Academic Excellence Award 2018 from IIT Hyderabad.
  • Academic Excellence Award 2016 from IIT Hyderabad.
  • National Talent Search Scheme (NTS) Scholarship 2013 from National Council of Educational Research and Training, India.

Team Competitions

Contact

  • sukrut.rao@mpi-inf.mpg.de
  • Max Planck Institute for Informatics
    Saarland Informatics Campus
    Campus E1 4, 66123 Saarbrücken, Germany