Search

Adversarial attacks pose significant threats to the reliability and safety of deep learning models, especially in critical domains such as medical imaging. This paper introduces a novel framework that integrates conformal prediction with game-theoretic defensive strategies to enhance model robustness against both known and unknown adversarial perturbations. We address three primary research questions: constructing valid and efficient conformal prediction sets under known attacks (RQ1), ensuring coverage under unknown attacks through conservative thresholding (RQ2), and determining optimal defensive strategies within a zero-sum game framework (RQ3). Our methodology involves training specialized defensive models against specific attack types and employing maximum and minimum classifiers to aggregate defenses effectively. Extensive experiments conducted on the MedMNIST datasets, including PathMNIST, OrganAMNIST, and TissueMNIST, demonstrate that our approach maintains high coverage guarantees while minimizing prediction set sizes. The game-theoretic analysis reveals that the optimal defensive strategy often converges to a singular robust model, outperforming uniform and simple strategies across all evaluated datasets. This work advances the state-of-the-art in uncertainty quantification and adversarial robustness, providing a reliable mechanism for deploying deep learning models in adversarial environments.

Dense Retrieval (DR) is now considered as a promising tool to enhance the memorization capacity of Large Language Models (LLM) such as GPT3 and GPT-4 by incorporating external memories. However, due to the paradigm discrepancy between text generation of LLM and DR, it is still an open challenge to integrate the retrieval and generation tasks in a shared LLM. In this paper, we propose an efficient LLM-Oriented Retrieval Tuner, namely LMORT, which decouples DR capacity from base LLM and non-invasively coordinates the optimally aligned and uniform layers of the LLM towards a unified DR space, achieving an efficient and effective DR without tuning the LLM itself. The extensive experiments on six BEIR datasets show that our approach could achieve competitive zero-shot retrieval performance compared to a range of strong DR models while maintaining the generation ability of LLM., Comment: 16 pages, 8 figures, 5 tables

We report observations of discrete charge states of a coherent dielectric two-level system (TLS) that is strongly coupled to an offset-charge-sensitive superconducting transmon qubit. We measure an offset charge of 0.072$e$ associated with the two TLS eigenstates, which have a transition frequency of 2.9 GHz and a relaxation time exceeding 3 ms. Combining measurements in the strong dispersive and resonant regime, we quantify both transverse and longitudinal couplings of the TLS-qubit interaction. We further perform joint tracking of TLS transitions and quasiparticle tunneling dynamics but find no intrinsic correlations. This study demonstrates microwave-frequency TLS as a source of low-frequency charge noise., Comment: 5 pages, 5 figures + 10 pages, 8 figures, 2 tables

Liquid holdup and mass transfer area are critical parameters for packed column design and CO2 capture efficiency prediction. In this paper, a framework was established for modeling the liquid-gas countercurrent flow hydrodynamics in a random packed column with pall rings. Besides the column-averaged information, the radial pall ring distribution, velocity, and liquid holdup profiles are obtained to study the entrance effect and the wall influence in the packed column. With local CFD data, the validated packing specific area ap and liquid velocity uL range for liquid holdup correlation is significantly expanded with respect to existing experimental or column-averaged CFD data. The proposed liquid holdup correlation $h_L \propto u_L^{0.44}$ indicates the random packed column falls in a viscous to turbulent transition regime and it covers a Reynolds Number range of [6.7-40.2]. The derived liquid holdup correlation is in good agreement with existing correlations developed using the column-averaged experimental data.

Few-shot dense retrieval (DR) aims to effectively generalize to novel search scenarios by learning a few samples. Despite its importance, there is little study on specialized datasets and standardized evaluation protocols. As a result, current methods often resort to random sampling from supervised datasets to create "few-data" setups and employ inconsistent training strategies during evaluations, which poses a challenge in accurately comparing recent progress. In this paper, we propose a customized FewDR dataset and a unified evaluation benchmark. Specifically, FewDR employs class-wise sampling to establish a standardized "few-shot" setting with finely-defined classes, reducing variability in multiple sampling rounds. Moreover, the dataset is disjointed into base and novel classes, allowing DR models to be continuously trained on ample data from base classes and a few samples in novel classes. This benchmark eliminates the risk of novel class leakage, providing a reliable estimation of the DR model's few-shot ability. Our extensive empirical results reveal that current state-of-the-art DR models still face challenges in the standard few-shot scene. Our code and data will be open-sourced at https://github.com/OpenMatch/ANCE-Tele., Comment: Work in progress

Data-driven modeling approaches can produce fast surrogates to study large-scale physics problems. Among them, graph neural networks (GNNs) that operate on mesh-based data are desirable because they possess inductive biases that promote physical faithfulness, but hardware limitations have precluded their application to large computational domains. We show that it is \textit{possible} to train a class of GNN surrogates on 3D meshes. We scale MeshGraphNets (MGN), a subclass of GNNs for mesh-based physics modeling, via our domain decomposition approach to facilitate training that is mathematically equivalent to training on the whole domain under certain conditions. With this, we were able to train MGN on meshes with \textit{millions} of nodes to generate computational fluid dynamics (CFD) simulations. Furthermore, we show how to enhance MGN via higher-order numerical integration, which can reduce MGN's error and training time. We validated our methods on an accompanying dataset of 3D $\text{CO}_2$-capture CFD simulations on a 3.1M-node mesh. This work presents a practical path to scaling MGN for real-world applications., Comment: ICLR 2023 Workshop on Physics for Machine Learning

This paper presents a distributed data-driven predictive control (DDPC) approach using the behavioral framework. It aims to design a network of controllers for an interconnected system with linear time-invariant (LTI) subsystems such that a given global (network-wide) cost function is minimized while desired control performance (e.g., network stability and disturbance rejection) is achieved using dissipativity in the quadratic difference form (QdF). By viewing dissipativity as a behavior and integrating it into the control design as a virtual dynamical system, the proposed approach carries out the entire design process in a unified framework with a set-theoretic viewpoint. This leads to an effective data-driven distributed control design, where the global design goal can be achieved by distributed optimization based on the local QdF conditions. The approach is illustrated by an example throughout the paper.

Computational countercurrent flow investigation in the structured packed column is a multiscale problem. Multiphase flow studies using volume of fluid (VOF) method in the representative elementary unit (REU) of the packed column can insight into the local hydrodynamics such as interfacial area, film thickness, etc. The interfacial area dictates the mass transfer in absorption process and thereby overall efficiency of column. Impacts of solvent's physical properties, liquid loads and static contact angle (SCA) on the interfacial area were examined earlier. In the present study, the dynamic contact angle (DCA) was used to explore the impact of contact angle hysteresis on the interfacial area. DCA has more pronounced impact on the interfacial area (10%) for aqueous solvent of 0.10M Sodium hydroxide (NaOH). The interfacial area shows undulation and does not achieve the pseudo-steady state. In contrary, the interfacial area gets a net pseudo-steady value for the aqueous solvent having 40% monoethanolamine (MEA) by weight. The wetting hysteresis was also explored via simulations conducted with initially dry and wetted sheets. For 0.10M NaOH aqueous solvent, the initially wetted sheets lead to slightly higher value of the interfacial area (10%) as compared to the initially dry sheets at the same liquid load and DCA. As expected, wetting hysteresis reduces with increasing liquid loads. On the other hand, wetting hysteresis is not significant for 40% MEA aqueous solvent which might be lower surface tension and higher viscosity. Overall, the effect of the dynamic contact angle is not pronounced as compared to those found in a flat surface.

Focus scanning is critically important in many terahertz (THz) imaging and sensing applications. A traditional single focusing transmitarray can achieve a good focus when the source is on-axis but moving the source off-axis produces a significant aberration. This paper presents a novel approach to reducing coma in off-axis scanning in the THz region. Here, a dual transmitarray solution is proposed, in which a transmitarray with an optimized phase profile is placed behind a regular phase profile transmitarray. A linearly polarized, dual-transmitarray antenna was fabricated for validation, and the focusing performances were experimentally characterized. The measured results are in good agreement with the theoretical ones. The generated spot of the dual-transmitarray antenna remains focused on an angle up to 50deg, with a -3 dB spot size of less than 4 mm at 290 GHz. The measured near-field sidelobes are all below -10 dB within the whole scanning range., Comment: 9 pages, 14 figures

In this paper, we investigate the instability in the standard dense retrieval training, which iterates between model training and hard negative selection using the being-trained model. We show the catastrophic forgetting phenomena behind the training instability, where models learn and forget different negative groups during training iterations. We then propose ANCE-Tele, which accumulates momentum negatives from past iterations and approximates future iterations using lookahead negatives, as "teleportations" along the time axis to smooth the learning process. On web search and OpenQA, ANCE-Tele outperforms previous state-of-the-art systems of similar size, eliminates the dependency on sparse retrieval negatives, and is competitive among systems using significantly more (50x) parameters. Our analysis demonstrates that teleportation negatives reduce catastrophic forgetting and improve convergence speed for dense retrieval training. Our code is available at https://github.com/OpenMatch/ANCE-Tele., Comment: Accepted to EMNLP 2022 main conference