Is this you? Claim this profile to correct it, add a bio and choose the work people see first.
Claim this profileAcademic lineage
Possible advisorsa guess from early papers, not confirmed
- Suggested from co-authorship
- Peter RichtárikSuggested from co-authorship
Is this you? Claim this profile to confirm or dismiss it.
Works91 from public data
- TrackingNet: A Large-Scale Dataset and Benchmark for Object Tracking in the Wild1,171
This work presents TrackingNet, the first large-scale dataset and benchmark for object tracking in the wild, which covers a wide selection of object classes in broad and diverse context and provides an extensive benchmark on TrackingNet by evaluating more than 20 trackers.
- The Visual Object Tracking VOT2015 Challenge Results902
The Visual Object Tracking challenge 2015, VOT2015, aims at comparing short-term single-object visual trackers that do not apply pre-learned models of object appearance and presents a new VOT 2015 dataset twice as large as in VOT2014 with full annotation of targets by rotated bounding boxes and per-frame attribute.
- 184
- Target Response Adaptation for Correlation Filter Tracking163
This work proposes a generic framework that can adaptively change the target response from frame to frame, so that the tracker is less sensitive to the cases where circular shifts do not reliably approximate translations.
- 138
- Computationally Budgeted Continual Learning: What Does Matter?95
It is found that, under compute-constrained setting, traditional CL approaches, with no exception, fail to outperform a simple minimal baseline that samples uniformly from memory, suggesting that most existing CL methods are particularly too computationally expensive for realistic budgeted deployment.
- Real-Time Evaluation in Online Continual Learning: A New Hope77
It is shown that a simple baseline outperforms state-of-the-art CL methods under this evaluation, questioning the applicability of existing methods in realistic settings and suggesting that the majority of existing CL literature is tailored to a specific class of streams that is not practical.
- 75
- SynthCLIP: Are We Ready for a Fully Synthetic CLIP Training?73
This work provides an analysis on CLIP models trained on synthetic data, providing insights on the data generation strategy, number of samples required, scaling trends, and resulting properties.
- Open Problems in Machine Unlearning for AI Safety63
This paper identifies key limitations that prevent unlearning from serving as a comprehensive solution for AI safety, particularly in managing dual-use knowledge in sensitive domains like cybersecurity and chemical, biological, radiological, and nuclear safety.
- When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitations59
It is shown that despite the continuous embedding space being more expressive than the discrete token space, soft-prompting and prefix-tuning are potentially less expressive than full fine-tuning, even with the same number of learnable parameters.
- 48
- Combating Adversaries with Anti-adversaries43
The anti-adversary layer is proposed, which generates an input perturbation in the opposite direction of the adversarial one and feeds the classifier a perturbed version of the input.
- 43
- 40
- On the Decision Boundaries of Neural Networks: A Tropical Geometry Perspective38
The main finding is that the decision boundaries of a simple network of the form (Affine, ReLU, Affine) are a subset of a tropical hypersurface, which is intimately related to a polytope formed by the convex hull of two zonotopes.
- Bi-Factorial Preference Optimization: Balancing Safety-Helpfulness in Language Models34
A supervised learning framework called Bi-Factorial Preference Optimization is proposed, which re-parameterizes a joint RLHF objective of both safety and helpfulness into a single supervised learning objective and significantly outperforms existing approaches in both safety and helpfulness.
- Mixture of Experts Made Intrinsically Interpretable30
MoE-X, a Mixture-of-Experts (MoE) language model designed to be interpretable, achieves performance comparable to dense models while significantly improving interpretability, with interpretability surpassing even sparse autoencoder (SAE)-based approaches.
- ToolTweak: An Attack on Tool Selection in LLM-based Agents29
This paper presents ToolTweak, a lightweight automatic attack that increases selection rates from a baseline of around 20% to as high as 81%, with strong transferability between open-source and closed-source models and evaluation of two defenses: paraphrasing and perplexity filtering.
- Prompting a Pretrained Transformer Can Be a Universal Approximator29
It is demonstrated that much smaller pretrained models than previously thought can be universal approximators when prefixed and offer Jackson-type bounds on the length of the prefix needed to approximate a function to a desired precision.
- 29
- 28
- Gabor Layers Enhance Network Robustness27
It is observed that architectures enhanced with Gabor layers gain a consistent boost in robustness over regular models and preserve high generalizing test performance, even though these layers come at a negligible increase in the number of parameters.
- Near to Mid-term Risks and Opportunities of Open-Source Generative AI23
It is argued for the responsible open sourcing of generative AI models in the near and medium term and differential benefits and risks of open versus closed source AI and present potential risk mitigation, ranging from best practices to calls for technical and scientific contributions.
- Rapid Adaptation in Online Continual Learning: Are We Evaluating It Right?23
This work proposes a novel metric for measuring adaptation of Online Continual Learning algorithms based on the accuracy on the near-future samples, where spurious correlations are removed, and finds that better generalization can be achieved by retaining and reusing past seen information.
- DeformRS: Certifying Input Deformations with Randomized Smoothing23
This work reformulate certification in randomized smoothing setting for both general vector field and parameterized deformations and proposes DeformRS-VF and Deform RS-Par, respectively, which scales to large networks on large input datasets.
- 22
- Beyond Linear Probes: Dynamic Safety Monitoring for Language Models19
Tuncated Polynomial Classifiers are introduced, a natural extension of linear probes for dynamic activation monitoring that compete with or outperform MLP-based probe baselines of the same size for harmful prompt classification, all the while being more interpretable than their black-box counterparts.
- Rethinking Safety in LLM Fine-tuning: An Optimization Perspective17
This work proposes a simple exponential moving average (EMA) momentum technique in parameter space that preserves safety performance by creating a stable optimization path and retains the original pre-trained model's safety properties, and proposes practical guidelines for maintaining both model performance and safety during adaptation.
- A Stochastic Derivative-Free Optimization Method with Importance Sampling: Theory and Learning to Control17
This paper proposes the first derivative free optimization method with importance sampling and derive new improved complexity results on non-convex, convex and strongly convex functions and test the method on a collection of continuous control tasks on MuJoCo environments with varying difficulty.
- A Stochastic Derivative-Free Optimization Method with Importance Sampling17
This paper proposes the first derivative free optimization method with importance sampling and derive new improved complexity results on non-convex, convex and strongly convex functions and test the method on a collection of continuous control tasks on MuJoCo environments with varying difficulty.
- Unforgotten Safety: Preserving Safety Alignment of Large Language Models with Continual Learning16
This paper considers the fine-tuning-as-a-service setup where the user uploads their data to a service provider to get a customized model that excels on the user's selected task.
- Don’t FREAK Out: A Frequency-Inspired Approach to Detecting Backdoor Poisoned Samples in DNNs15
This work investigates the frequency sensitivity of Deep Neural Networks when presented with clean samples versus poisoned samples and proposes FREAK, a frequency-based poisoned sample detection algorithm that is simple yet effective against frequency backdoor attacks.
- MIP against Agent: Malicious Image Patches Hijacking Multimodal OS Agents13
It is shown that MIPs generalise across user prompts and screen configurations, and that they can hijack multiple OS agents even during the execution of benign instructions, which exposes critical security vulnerabilities in OS agents that have to be carefully addressed before their widespread deployment.
- Label Delay in Online Continual Learning13
A new continual learning framework with explicit modeling of the label delay between data and label streams over time steps is introduced, showing experimentally that the method is the least affected by the labeldelay factor and in some cases successfully recovers the accuracy of the non-delayed counterpart.
- Efficient Lifelong Model Evaluation in an Era of Rapid Progress11
An efficient framework for model evaluation, Sort&Search (S&S), is introduced, which reuses previously evaluated models by leveraging dynamic programming algorithms to selectively rank and sub-select test samples and achieves highly-efficient approximate accuracy measurement.
- 10
- Towards Interpretable Deep Local Learning with Successive Gradient Reconciliation9
This paper theoretically study non-greedy layer-wise training and shows that the convergence cannot be assured when the local gradient in a module w.r.t. its input is not reconciled with the local gradient in the previous module w.r.t. its output.
- 9
- Constrained Clustering: General Pairwise and Cardinality Constraints9
The experimental results show that the proposed method exploits the constraints to achieve perfect clustering performance with improved clustering to $2-5$ % in classical clustering metrics, e.g., Adjusted Random Index (ARI), Mirkin's Index (MI), and Huber’s Index (HI), outerperfomring all compared-againts methods across the board.
- Shh, don't say that! Domain Certification in LLMs8
This work introduces domain certification; a guarantee that accurately characterizes the out-of-domain behavior of language models that provides adversarial bounds as a certificate and proposes a simple yet effective approach, which is called VALID, that provides adversarial bounds as a certificate.
- Continual Learning on a Diet: Learning from Sparsely Labeled Streams Under Constrained Computation8
DietCL outperforms, by a large margin, all existing supervised CL algorithms as well as more recent continual semi-supervised methods, and meticulously allocates computational budget for both unlabeled and labeled data jointly.
- 7
- 7
- Towards Certification of Uncertainty Calibration under Adversarial Attacks7
It is shown that attacks can significantly harm calibration, and thus certified calibration is proposed as worst-case bounds on calibration under adversarial perturbations.
- SimCS: Simulation for Domain Incremental Online Continual Segmentation7
This work explores the problem of Online Domain-Incremental Continual Segmentation (ODICS), where the model is continually trained over batches of densely labeled images from different domains, with limited computation and no information about the task boundaries.
- 7
- 6
- 6
- The Authorization-Execution Gap Is a Major Safety and Security Problem in Open-World Agents5
It is argued that many observed agent failures can be traced to three structural sources of AEG: delegation-level incompleteness, channel-level corruption, and composition-level fragmentation, and that papers on open-world agents should report not only outcome-level metrics such as task success or attack resistance, but also process-level evidence showing where AEG was detected, constrained, and attributed to a structural source during execution.
- The Alignment Curse: Modality Alignment Supercharges Audio Attacks via Text Transfer5
This work analyzes the connection between modality alignment and cross-modality jailbreak transfer, and demonstrates strong cross-model transferability and shows that text-transferred audio attacks remain effective even under a stricter audio-only access threat model.
- 5
- 5
- 5
- Constrained K-means with General Pairwise and Cardinality Constraints5
This paper enforces pairwise and cardinality constraints into a unified clustering model starting with the integer program formulation of the standard K-means and derives an equivalent continuous reformulation with simple constraints, which can be efficiently solved by Alternating Direction Method of Multipliers (ADMM).
- 4
- From Categories to Classifiers: Name-Only Continual Learning by Exploring the Web3
Overall, this paper underscores the potential of using uncurated webly-supervised data to mitigate the challenges associated with manual data labeling in continual learning.
- Certifying Ensembles: A General Certification Theory with S-Lipschitzness3
This work generalise Lipschitz continuity by introducing S-Lipschitzer classifiers, which are used to analyse the theoretical robustness of ensembles and results are precise conditions when ensemble of robust classifiers are more robust than any constituent classifier.
- Illusory Attacks: Information-Theoretic Detectability Matters in Adversarial Attacks3
It is concluded that agents must seek realism feedback to be robust to illusionary attacks, a novel form of attack that is consistent with the world model of the victim agent.
- 3
- ClusTR: Clustering Training for Robustness.3
ClusTR (Clustering Training for Robustness), a clustering-based and adversary-free training framework to learn robust models, is proposed and can be equipped with simple and fast adversarial training to improve the current state-of-the-art in robustness.
- Robust Gabor Networks3
The effect of replacing the first layers of neural network architectures with convolutional layers that are based on Gabor filters with learnable parameters is explored, and architectures enhanced with Gabor layers gain advantages in terms of robustness when compared to the regular models.
- 3
- 3
- 2
- 2
- Reliable and Responsible Foundation Models: A Comprehensive Survey1
Critical issues, including bias and fairness, security and privacy, uncertainty, explainability, and distribution shift are explored, as well as methods like alignment and Artificial Intelligence-Generated Content (AIGC) detection.
- 1
- 1
- 1
- 1
- Rethinking Clustering for Robustness1
This paper studies how encouraging semantically-aligned features during deep neural network training can increase network robustness, and proposes ClusTR (Clustering Training for Robustness), a clustering-based and adversary-free training framework to learn robust models.
- 1
- 1
- –
- –
- –
- –
- –
- $D^2$-Monitor: Dynamic Safety Monitoring for Diffusion LLMs via Hesitation Signals–
This work proposes D 2 -Monitor, a bi-level safety monitor for D-LLMs, which achieves state-of-the-art performance with a compact parameter footprint, and exhibits the best trade-off between effectiveness and efficiency relative to 8 baselines.
- –
- –
- –
- –
- –
- –
- –
- –
- Expected Tight Bounds for Robust Deep Neural Network Training–
This work was supported by the King Abdullah University of Science and Technology (KAUST) Office of Sponsored Research.
- Probabilistically True and Tight Bounds for Robust Deep Neural Network Training.–
With such tight bounds, it is demonstrated that a simple standard training procedure can achieve the best robustness-accuracy trade-off across several architectures on both MNIST and CIFAR10.
- –
Publication data from OpenAlex, with missing venues and authors filled in from Crossref; citation counts are the higher of OpenAlex and Semantic Scholar, last synced 2026-10-13. One-sentence summaries under some papers are written by Semantic Scholar’s model. Citation counts may be lower than on Google Scholar, which indexes more sources.
Report an error
Wrong papers, two people merged into one, or a profile that should not be here? Tell us and we will fix or hide it.