Search

A plethora of jailbreaking attacks have been proposed to obtain harmful responses from safety-tuned LLMs. In their original settings, these methods all largely succeed in coercing the target output, but their attacks vary substantially in fluency and computational effort. In this work, we propose a unified threat model for the principled comparison of these methods. Our threat model combines constraints in perplexity, measuring how far a jailbreak deviates from natural text, and computational budget, in total FLOPs. For the former, we build an N-gram model on 1T tokens, which, in contrast to model-based perplexity, allows for an LLM-agnostic and inherently interpretable evaluation. We adapt popular attacks to this new, realistic threat model, with which we, for the first time, benchmark these attacks on equal footing. After a rigorous comparison, we not only find attack success rates against safety-tuned modern models to be lower than previously presented but also find that attacks based on discrete optimization significantly outperform recent LLM-based attacks. Being inherently interpretable, our threat model allows for a comprehensive analysis and comparison of jailbreak attacks. We find that effective attacks exploit and abuse infrequent N-grams, either selecting N-grams absent from real-world text or rare ones, e.g. specific to code datasets.

Osteoarthritis (OA) is the most common musculoskeletal disease, which has no cure. Knee OA (KOA) is one of the highest causes of disability worldwide, and it costs billions of United States dollars to the global community. Prediction of KOA progression has been of high interest to the community for years, as it can advance treatment development through more efficient clinical trials and improve patient outcomes through more efficient healthcare utilization. Existing approaches for predicting KOA, however, are predominantly static, i.e. consider data from a single time point to predict progression many years into the future, and knee level, i.e. consider progression in a single joint only. Due to these and related reasons, these methods fail to deliver the level of predictive performance, which is sufficient to result in cost savings and better patient outcomes. Collecting extensive data from all patients on a regular basis could address the issue, but it is limited by the high cost at a population level. In this work, we propose to go beyond static prediction models in OA, and bring a novel Active Sensing (AS) approach, designed to dynamically follow up patients with the objective of maximizing the number of informative data acquisitions, while minimizing their total cost over a period of time. Our approach is based on Reinforcement Learning (RL), and it leverages a novel reward function designed specifically for AS of disease progression in more than one part of a human body. Our method is end-to-end, relies on multi-modal Deep Learning, and requires no human input at inference time. Throughout an exhaustive experimental evaluation, we show that using RL can provide a higher monetary benefit when compared to state-of-the-art baselines.

Superconducting qubits based on Al/AlOx/Al Josephson junction are one of the most promising candidates for the physical implementation of universal quantum computers. Due to scalability and compatibility with the state-of-the-art nanoelectronic processes one can fabricate hundreds of qubits on a single silicon chip. However, decoherence in these systems caused by two-level-systems in amorphous dielectrics, including a tunneling barrier AlOx, is one of the major problems. We report on a Josephson junction thermal annealing process development to crystallize an amorphous barrier oxide (AlOx). The dependences of the thermal annealing parameters on the room temperature resistance are obtained. The developed method allows not only to increase the Josephson junction resistance by 175%, but also to decrease by 60% with precisions of 10% in Rn. Finally, theoretical assumptions about the structure modification in tunnel barrier are proposed. The suggested thermal annealing approach can be used to form a stable and reproducible tunneling barriers and scalable frequency trimming for a widely used fixed-frequency transmon qubits.

One of the practical limitations of solid-state superconducting quantum processors technology is frequency crowding due to low qubits fabrication reproducibility. Josephson junction 100 nm-scale nonlinear inductance of the qubits still suffers from Dolan-bridge shadow evaporation process. Here, we report on a robust wafer-scale Al/AlOx/Al Dolan-bridge Josephson junction (JJ) process using preliminary shadow evaporation bias resist mask correction and comprehensive oxidation optimization. We introduce topology correction model for two-layer resist mask biasing at a wafer-scale, which takes into account an evaporation source geometry. It results in Josephson junction area variation coefficient improvement down to 1.1% for the critical dimensions from 130x170 nm2 to 130x670 nm2 over 70x70 mm2 (49 cm2) wafer working area. Next, we investigate JJ oxidation process (oxidation method, pressure and time) and its impact on a room temperature resistance reproducibility. Finally, we combine both shadow evaporation bias correction and oxidation best practices for 4-inch wafers improving room temperature resistance variation coefficient down to 6.0/5.2/4.1% for 0.025 {\mu}m2 JJ area and 4.0/3.4/2.3% for 0.090 {\mu}m2 JJ area for 49/25/16 cm2 wafer working area correspondingly. The proposed model and oxidation method can be useful for robust wafer-scale superconducting quantum processors fabrication.

Learning representations that generalize to novel compositions of known concepts is crucial for bridging the gap between human and machine perception. One prominent effort is learning object-centric representations, which are widely conjectured to enable compositional generalization. Yet, it remains unclear when this conjecture will be true, as a principled theoretical or empirical understanding of compositional generalization is lacking. In this work, we investigate when compositional generalization is guaranteed for object-centric representations through the lens of identifiability theory. We show that autoencoders that satisfy structural assumptions on the decoder and enforce encoder-decoder consistency will learn object-centric representations that provably generalize compositionally. We validate our theoretical result and highlight the practical relevance of our assumptions through experiments on synthetic image data., Comment: The first four authors contributed equally

Knee Osteoarthritis (KOA) is a highly prevalent chronic musculoskeletal condition with no currently available treatment. The manifestation of KOA is heterogeneous and prediction of its progression is challenging. Current literature suggests that the use of multi-modal data and advanced modeling methods, such as the ones based on Deep Learning, has promise in tackling this challenge. To date, however, the evidence on the efficacy of this approach is limited. In this study, we leveraged recent advances in Deep Learning and, using a Transformer approach, developed a unified framework for the multi-modal fusion of knee imaging data. Subsequently, we analyzed its performance across a range of scenarios by investigating multiple progression horizons -- from short-term to long-term. We report our findings using a large cohort (n=2421-3967) derived from the Osteoarthritis Initiative dataset. We show that structural knee MRI allows identifying radiographic KOA progressors on par with multi-modal fusion approaches, achieving an area under the ROC curve (ROC AUC) of 0.70-0.76 and Average Precision (AP) of 0.15-0.54 in 2-8 year horizons. Progression within 1 year was better predicted with a multi-modal method using X-ray, structural, and compositional MR images -- ROC AUC of 0.76(0.04), AP of 0.13(0.04) -- or via clinical data. Our follow-up analysis generally shows that prediction from the imaging data is more accurate for post-traumatic subjects, and we further investigate which subject subgroups may benefit the most. The present study provides novel insights into multi-modal imaging of KOA and brings a unified data-driven framework for studying its progression in an end-to-end manner, providing new tools for the design of more efficient clinical trials. The source code of our framework and the pre-trained models are made publicly available.

Excitable media are ubiquitous in nature, and in such systems the local excitation tends to self-organize in traveling waves, or in rotating spiral-shaped patterns in two or three spatial dimensions. Examples include waves during a pandemic or electrical scroll waves in the heart. Here we show that such phenomena can be extended to a space of four or more dimensions and propose that connections of excitable elements in a network setting can be regarded as additional spatial dimensions. Numerical simulations are performed in four dimensions using the FitzHugh-Nagumo model, showing that the vortices rotate around a two-dimensional surface which we define as the superfilament. Evolution equations are derived for general superfilaments of co-dimension two in an N -dimensional space and their equilibrium configurations are proven to be minimal surfaces. We suggest that biological excitable systems, such as the heart or brain which have non-local connections can be regarded, at least partially, as multidimensional excitable media and discuss further possible studies in this direction.

Adversarial examples are transferable between different models. In our paper, we propose to use this property for multi-step domain adaptation. In unsupervised domain adaptation settings, we demonstrate that replacing the source domain with adversarial examples to $\mathcal{H} \Delta \mathcal{H}$-divergence can improve source classifier accuracy on the target domain. Our method can be connected to most domain adaptation techniques. We conducted a range of experiments and achieved improvement in accuracy on Digits and Office-Home datasets.

It was shown that adversarial examples improve object recognition. But what about their opposite side, easy examples? Easy examples are samples that the machine learning model classifies correctly with high confidence. In our paper, we are making the first step toward exploring the potential benefits of using easy examples in the training procedure of neural networks. We propose to use an auxiliary batch normalization for easy examples for the standard and robust accuracy improvement.

Many physical, chemical and biological processes rely on intrinsic oscillations to employ resonance responses to external stimuli of certain frequency. Such resonance phenomena in biological systems are typically explained by one of two mechanisms: either a classical linear resonance of harmonic oscillator, or entrainment and phase locking of nonlinear limit cycle oscillators subjected to periodic forcing. Here, we discover a nonlinear mechanism, which does not require intrinsic oscillations. Instead, the resonant frequency dependence arises from coupling between an excitable and a monostable region of the medium. This composite system is endowed with emergent bistability between a stable steady state and stable spatiotemporal oscillations. The resonant transition from stable state to oscillatory state is induced by waves of particular frequency travelling through the medium. This transition to the spatiotemporal oscillatory state requires accumulation of multiple waves, resulting in the exclusion of lower frequencies. The cutting off of high frequencies is realized by damping of wave amplitude in the monostable zone and then by activating amplitude sensitive dynamics in the monostable units. We demonstrate this new resonance mechanism in a simplistic reaction-diffusion model. Also, we reveal this collective resonance mechanism in in-vitro experiments and detailed biophysical simulations representing a major type of arrhythmia. We further demonstrate, both experimentally and theoretically, that the ongoing spatiotemporal oscillations, such as ectopic activity in cardiac tissue, can be stopped by travelling waves of high frequency. Overall, we claim the universality of this resonance mechanism in a broad class of nonlinear biophysical systems. Specifically, we hypothesize that such phenomena could be found in neuronal systems as an alternative to traditional resonant processes.

We present a novel algorithm for domain adaptation using optimal transport. In domain adaptation, the goal is to adapt a classifier trained on the source domain samples to the target domain. In our method, we use optimal transport to map target samples to the domain named source fiction. This domain differs from the source but is accurately classified by the source domain classifier. Our main idea is to generate a source fiction by c-cyclically monotone transformation over the target domain. If samples with the same labels in two domains are c-cyclically monotone, the optimal transport map between these domains preserves the class-wise structure, which is the main goal of domain adaptation. To generate a source fiction domain, we propose an algorithm that is based on our finding that adversarial attacks are a c-cyclically monotone transformation of the dataset. We conduct experiments on Digits and Modern Office-31 datasets and achieve improvement in performance for simple discrete optimal transport solvers for all adaptation tasks.