Search

Evaluating the safety of AI Systems is a pressing concern for organizations deploying them. In addition to the societal damage done by the lack of fairness of those systems, deployers are concerned about the legal repercussions and the reputational damage incurred by the use of models that are unsafe. Safety covers both what a model does; e.g., can it be used to reveal personal information from its training set, and how a model was built; e.g., was it only trained on licensed data sets. Determining the safety of an AI system requires gathering information from a wide set of heterogeneous sources including safety benchmarks and technical documentation for the set of models used in that system. In addition, responsible use is encouraged through mechanisms that advise and help the user to take mitigating actions where safety risks are detected. We present Usage Governance Advisor which creates semi-structured governance information, identifies and prioritizes risks according to the intended use case, recommends appropriate benchmarks and risk assessments and importantly proposes mitigation strategies and actions., Comment: 9 pages, 8 figures, AAAI workshop submission

Quantum computing is an emerging technology whose positive and negative impacts on society are not yet fully known. As government, individuals, institutions, and corporations fund and develop this technology, they must ensure that they anticipate its impacts, prepare for its consequences, and steer its development in such a way that it enables the most good and prevents the most harm. However, individual stakeholders are not equipped to fully anticipate these consequences on their own it requires a diverse community that is well-informed about quantum computing and its impacts. Collaborations and community-building across domains incorporating a variety of viewpoints, especially those from stakeholders most likely to be harmed, are fundamental pillars of developing and deploying quantum computing responsibly. This paper reviews responsible quantum computing proposals and literature, highlights the challenges in implementing these, and presents strategies developed at IBM aimed at building a diverse community of users and stakeholders to support the responsible development of this technology.

Pixel-based language models have emerged as a compelling alternative to subword-based language modelling, particularly because they can represent virtually any script. PIXEL, a canonical example of such a model, is a vision transformer that has been pre-trained on rendered text. While PIXEL has shown promising cross-script transfer abilities and robustness to orthographic perturbations, it falls short of outperforming monolingual subword counterparts like BERT in most other contexts. This discrepancy raises questions about the amount of linguistic knowledge learnt by these models and whether their performance in language tasks stems more from their visual capabilities than their linguistic ones. To explore this, we probe PIXEL using a variety of linguistic and visual tasks to assess its position on the vision-to-language spectrum. Our findings reveal a substantial gap between the model's visual and linguistic understanding. The lower layers of PIXEL predominantly capture superficial visual features, whereas the higher layers gradually learn more syntactic and semantic abstractions. Additionally, we examine variants of PIXEL trained with different text rendering strategies, discovering that introducing certain orthographic constraints at the input level can facilitate earlier learning of surface-level features. With this study, we hope to provide insights that aid the further development of pixel-based language models., Comment: 9 pages, Accepted to EMNLP 2025 Main

We consider the problem of single link failure in an elastic optical network, (also known as flex-grid WDM network). The task is to reroute optical connections that go through the broken link using free capacity of other links of the network. Nowadays, dynamic restoration gains popularity, in which the possiblity of rerouting is only inspected after a link failure is detected. Since the problem of recovery is NP-hard, heuristic algorithms are used to either find such routes, or suggest that the routes do not exist. In order to understand the quality of these heuristics, often mixed integer linear programming is used to obtain exact positive and negative answers. We present a detailed such model that checks whether restoration is possible without the use of additional regenerators. This means, that the new light paths need to satisfy a length constraint. As preprossing we apply a trimming procedure that takes advantage of this length constraint, and significantly speeds up the evaluation of these models. Our model is more general, and besides solving the problem of link restoration, also solves the full problem of wavelength and spectrum assignment., Comment: Accepted for IEEE BlackSeaCom conference, 7 pages

The elective clinical target volume (CTV-N) in oropharyngeal squamous cell carcinoma (OPSCC) is currently based mostly on the prevalence of lymph node metastases in different lymph node levels (LNLs) for a given primary tumor location. We present a probabilistic model for ipsilateral lymphatic spread that can quantify the microscopic nodal involvement risk based on an individual patient's T-category and clinical involvement of LNLs at diagnosis. We extend a previously published hidden Markov model (HMM), which models the LNLs (I, II, III, IV, V, and VII) as hidden binary random variables (RVs). Each represents a patient's true state of lymphatic involvement. Clinical involvement at diagnosis represents the observed binary RVs linked to the true state via sensitivity and specificity. The primary tumor and the hidden RVs are connected in a graph. Each edge represents the conditional probability of metastatic spread per abstract time-step, given disease at the edge's starting node. To learn these probabilities, we draw Markov chain Monte Carlo samples from the likelihood of a dataset (686 OPSCC patients) from three institutions. We compute the model evidence using thermodynamic integration for different graphs to determine which describes the data best. The graph maximizing the model evidence connects the tumor to each LNL and the LNLs I through V in order. It predicts the risk of occult disease in level IV is below 5% if level III is clinically negative, and that the risk of occult disease in level V is below 5% except for advanced T-category (T3 and T4) patients with clinical involvement of levels II, III, and IV. The provided statistical model of nodal involvement in OPSCC patients trained on multi-institutional data may guide the design of clinical trials on volume-deescalated treatment of OPSCC and contribute to more personal guidelines on elective nodal treatment., Comment: 17 pages, 12 figures, 7 tables, submitted to Physics in Medicine and Biology

Recent advances in quantum computers are demonstrating the ability to solve problems at a scale beyond brute force classical simulation. As such, a widespread interest in quantum algorithms has developed in many areas, with optimization being one of the most pronounced domains. Across computer science and physics, there are a number of different approaches for major classes of optimization problems, such as combinatorial optimization, convex optimization, non-convex optimization, and stochastic extensions. This work draws on multiple approaches to study quantum optimization. Provably exact versus heuristic settings are first explained using computational complexity theory - highlighting where quantum advantage is possible in each context. Then, the core building blocks for quantum optimization algorithms are outlined to subsequently define prominent problem classes and identify key open questions that, if answered, will advance the field. The effects of scaling relevant problems on noisy quantum devices are also outlined in detail, alongside meaningful benchmarking problems. We underscore the importance of benchmarking by proposing clear metrics to conduct appropriate comparisons with classical optimization techniques. Lastly, we highlight two domains - finance and sustainability - as rich sources of optimization problems that could be used to benchmark, and eventually validate, the potential real-world impact of quantum optimization., Comment: Updated title to match journal version

Context: Internet of Things (IoT) has become an important kind of distributed systems thanks to the wide-spread of cheap embedded devices equipped with different networking technologies. Although ubiquitous, developing IoT systems remains challenging. Inquiry: A recent field study with 194 IoT developers identifies debugging as one of the main challenges faced when developing IoT systems. This comes from the lack of debugging tools taking into account the unique properties of IoT systems such as non-deterministic data, and hardware restricted devices. On the one hand, offline debuggers allow developers to analyse post-failure recorded program information, but impose too much overhead on the devices while generating such information. Furthermore, the analysis process is also time-consuming and might miss contextual information relevant to find the root cause of bugs. On the other hand, online debuggers do allow debugging a program upon a failure while providing contextual information (e.g., stack trace). In particular, remote online debuggers enable debugging of devices without physical access to them. However, they experience debugging interference due to network delays which complicates bug reproducibility, and have limited support for dynamic software updates on remote devices. Approach: This paper proposes out-of-things debugging, an online debugging approach especially designed for IoT systems. The debugger is always-on as it ensures constant availability to for instance debug post-deployment situations. Upon a failure or breakpoint, out-of-things debugging moves the state of a deployed application to the developer's machine. Developers can then debug the application locally by applying operations (e.g., step commands) to the retrieved state. Once debugging is finished, developers can commit bug fixes to the device through live update capabilities. Finally, by means of a fine-grained flexible interface for accessing remote resources, developers have full control over the debugging overhead imposed on the device, and the access to device hardware resources (e.g., sensors) needed during local debugging. Knowledge: Out-of-things debugging maintains good properties of remote debugging as it does not require physical access to the device to debug it, while reducing debugging interference since there are no network delays on operations (e.g., stepping) issued on the debugger since those happen locally. Furthermore, device resources are only accessed when requested by the user which further mitigates overhead and opens avenues for mocking or simulation of non-accessed resources. Grounding: We implemented an out-of-things debugger as an extension to a WebAssembly Virtual Machine and benchmarked its suitability for IoT. In particular, we compared our solution to remote debugging alternatives based on metrics such as network overhead, memory usage, scalability, and usability in production settings. From the benchmarks, we conclude that our debugger exhibits competitive performance in addition to confining overhead without sacrificing debugging convenience and flexibility. Importance: Out-of-things debugging enables debugging of IoT systems by means of classical online operations (e.g., stepwise execution) while addressing IoT-specific concerns (e.g., hardware limitations). We show that having the debugger always-on does not have to come at cost of performance loss or increased overhead but instead can enforce a smooth-going and flexible debugging experience of IoT systems.

Nowadays, intrusion detection systems based on deep learning deliver state-of-the-art performance. However, recent research has shown that specially crafted perturbations, called adversarial examples, are capable of significantly reducing the performance of these intrusion detection systems. The objective of this paper is to design an efficient transfer learning-based adversarial detector and then to assess the effectiveness of using multiple strategically placed adversarial detectors compared to a single adversarial detector for intrusion detection systems. In our experiments, we implement existing state-of-the-art models for intrusion detection. We then attack those models with a set of chosen evasion attacks. In an attempt to detect those adversarial attacks, we design and implement multiple transfer learning-based adversarial detectors, each receiving a subset of the information passed through the IDS. By combining their respective decisions, we illustrate that combining multiple detectors can further improve the detectability of adversarial traffic compared to a single detector in the case of a parallel IDS design., Comment: This is a preprint of an already published journal paper

We study connections between expansion in bipartite graphs and efficient online matching modeled via several games. In the basic game, an opponent switches {\em on} and {\em off} nodes on the left side and, at any moment, at most $K$ nodes may be on. Each time a node is switched on, it must be irrevocably matched with one of its neighbors. A bipartite graph has $e$-expansion up to $K$ if every set $S$ of at most $K$ left nodes has at least $e\#S$ neighbors. If all left nodes have degree $D$ and $e$ is close to $D$, then the graph is a lossless expander. We show that lossless expanders allow for a polynomial time strategy in the above game, and, furthermore, with a slight modification, they allow a strategy running in time $O(D \log N)$, where $N$ is the number of left nodes. Using this game and a few related variants, we derive applications in data structures and switching networks. Namely, (a) 1-query bitprobe storage schemes for dynamic sets (previous schemes work only for static sets),(b) explicit space- and time-efficient storage schemes for static and dynamic sets with non-adaptive access to memory (the first fully dynamic dictionary with non-adaptive probing using almost optimal space), and (c) non-explicit constant depth non-blocking $N$-connectors with poly$(\log N)$ time path finding algorithms whose size is optimal within a factor of $O(\log N)$ (previous connectors are double-exponentially slower)., Comment: The exposition has been improved and a few minor issues have been fixed

While the potential for peer-to-peer electricity trading, where households trade surplus electricity with peers in a local energy market, is rapidly growing, the drivers of participation in this trading scheme have been understudied so far. In particular, there is a dearth of research on the role of non-monetary incentives for trading surplus electricity, despite their potentially important role. This paper presents the first discrete choice experiment conducted with prosumers (i.e. proactive households actively managing their electricity production and consumption) in the Netherlands. Electricity trading preferences are analyzed regarding economic, environmental, social and technological parameters, based on survey data (N = 74). The dimensions most valued by prosumers are the environmental and, to a lesser extent, economic dimensions, highlighting the key motivating roles of environmental factors. Furthermore, a majority of prosumers stated they would provide surplus electricity for free or for non-monetary compensations, especially to energy-poor households. These observed trends were more pronounced among members of energy cooperatives. This suggests that peer-to-peer energy trading can advance a socially just energy transition. Regarding policy recommendations, these findings point to the need for communicating environmental and economic benefits when marketing P2P electricity trading platforms and for technical designs enabling effortless and customizable transactions, Comment: 44 pages, 7 figures, submitted to Energy Policy