arXiv subjects
Publications and source records attributed to Sebastian Schmidt.
The Valida instruction set architecture is designed for implementation in zkVMs to optimize for fast, efficient execution proving. This specification intends to guide implementors of zkVMs and compiler toolchains for Valida. It provides an unambiguous definition of the semantics of Valida programs and may be used as a starting point for formalization efforts.
This work presents a novel flow-based molecular communication (MC) testbed using spectral sensing and ink intensity estimation to enable real-time multi-molecule (MUMO) transmission. MUMO communication opens up crucial opportunities for increased throughput as well as implementing more complex coding, modulation, and resource allocation strategies for MC testbeds. An estimator using non-invasive spectral sensing at the receiver is proposed based on a simple absorption model. We conduct in-depth channel impulse response (CIR) measurements and a preliminary communication performance evaluation. Additionally, a simple analytical model is used to check the consistency of the CIRs. The results indicate that by utilizing MUMO transmission, on-off-keying, and a simple difference detector, the testbed can achieve up to 3 bits per second for near-error-free communication, which is on par with comparable testbeds that utilize more sophisticated coding or detection methods. Our platform lays the ground for implementing MUMO communication and evaluating various physical layer and networking techniques based on multiple molecule types in future MC testbeds in real time.
As the data demand for deep learning models increases, active learning (AL) becomes essential to strategically select samples for labeling, which maximizes data efficiency and reduces training costs. Real-world scenarios necessitate the consideration of incomplete data knowledge within AL. Prior works address handling out-of-distribution (OOD) data, while another research direction has focused on category discovery. However, a combined analysis of real-world considerations combining AL with out-of-distribution data and category discovery remains unexplored. To address this gap, we propose Joint Out-of-distribution filtering and data Discovery Active learning (Joda) , to uniquely address both challenges simultaneously by filtering out OOD data before selecting candidates for labeling. In contrast to previous methods, we deeply entangle the training procedure with filter and selection to construct a common feature space that aligns known and novel categories while separating OOD samples. Unlike previous works, Joda is highly efficient and completely omits auxiliary models and training access to the unlabeled pool for filtering or selection. In extensive experiments on 18 configurations and 3 metrics, \ours{} consistently achieves the highest accuracy with the best class discovery to OOD filtering balance compared to state-of-the-art competitor approaches.
Proposals for molecular communication networks as part of a future internet of bio-nano-things have become more intricate and the question of practical implementation is gaining more importance. One option is to apply detailed chemical modeling to capture more realistic effects of computing processes in biological systems. In this paper, we present ChemSICal, a detailed model for implementing the successive interference cancellation (SIC) algorithm for molecular multiple access in diffusion-based molecular communication networks as a chemical reaction network (CRN). We describe the structure of the model as a number of smaller reaction blocks, their speed controlled by reaction rate constants (RRCs). Deterministic and stochastic methods are utilized to first iteratively improve the choice of RRCs and subsequently investigate the performance of the model in terms of an error probability. We analyze the model's sensitivity to parameter changes and find that the analytically optimal values for the non-chemical model do not necessarily translate to the chemical domain. This necessitates careful optimization, especially of the RRCs, which are crucial for the successful operation of the ChemSICal system.
Information from social media can provide essential information for emergency response during natural disasters in near real-time. However, it is difficult to identify the disaster-related posts among the large amounts of unstructured data available. Previous methods often use keyword filtering, topic modelling or classification-based techniques to identify such posts. Active Learning (AL) presents a promising sub-field of Machine Learning (ML) that has not been used much in the field of text classification of social media content. This study therefore investigates the potential of AL for identifying disaster-related Tweets. We compare a keyword filtering approach, a RoBERTa model fine-tuned with generic data from CrisisLex, a base RoBERTa model trained with AL and a fine-tuned RoBERTa model trained with AL regarding classification performance. For testing, data from CrisisLex and manually labelled data from the 2021 flood in Germany and the 2023 Chile forest fires were considered. The results show that generic fine-tuning combined with 10 rounds of AL outperformed all other approaches. Consequently, a broadly applicable model for the identification of disaster-related Tweets could be trained with very little labelling effort. The model can be applied to use cases beyond this study and provides a useful tool for further research in social media analysis.
The trustworthiness of AI applications has been the subject of recent research and is also addressed in the EU's recently adopted AI Regulation. The currently emerging foundation models in the field of text, speech and image processing offer completely new possibilities for developing AI applications. This whitepaper shows how the trustworthiness of an AI application developed with foundation models can be evaluated and ensured. For this purpose, the application-specific, risk-based approach for testing and ensuring the trustworthiness of AI applications, as developed in the 'AI Assessment Catalog - Guideline for Trustworthy Artificial Intelligence' by Fraunhofer IAIS, is transferred to the context of foundation models. Special consideration is given to the fact that specific risks of foundation models can have an impact on the AI application and must also be taken into account when checking trustworthiness. Chapter 1 of the white paper explains the fundamental relationship between foundation models and AI applications based on them in terms of trustworthiness. Chapter 2 provides an introduction to the technical construction of foundation models and Chapter 3 shows how AI applications can be developed based on them. Chapter 4 provides an overview of the resulting risks regarding trustworthiness. Chapter 5 shows which requirements for AI applications and foundation models are to be expected according to the draft of the European Union's AI Regulation and Chapter 6 finally shows the system and procedure for meeting trustworthiness requirements.
Conversational search engines such as YouChat and Microsoft Copilot use large language models (LLMs) to generate responses to queries. It is only a small step to also let the same technology insert ads within the generated responses - instead of separately placing ads next to a response. Inserted ads would be reminiscent of native advertising and product placement, both of which are very effective forms of subtle and manipulative advertising. Considering the high computational costs associated with LLMs, for which providers need to develop sustainable business models, users of conversational search engines may very well be confronted with generated native ads in the near future. In this paper, we thus take a first step to investigate whether LLMs can also be used as a countermeasure, i.e., to block generated native ads. We compile the Webis Generated Native Ads 2024 dataset of queries and generated responses with automatically inserted ads, and evaluate whether LLMs or fine-tuned sentence transformers can detect the ads. In our experiments, the investigated LLMs struggle with the task but sentence transformers achieve precision and recall values above 0.9.
Active learning (AL) reduces the amount of labeled data needed to train a machine learning model by intelligently choosing which instances to label. Classic pool-based AL requires all data to be present in a datacenter, which can be challenging with the increasing amounts of data needed in deep learning. However, AL on mobile devices and robots, like autonomous cars, can filter the data from perception sensor streams before reaching the datacenter. We exploited the temporal properties for such image streams in our work and proposed the novel temporal predicted loss (TPL) method. To evaluate the stream-based setting properly, we introduced the GTA V streets and the A2D2 streets dataset and made both publicly available. Our experiments showed that our approach significantly improves the diversity of the selection while being an uncertainty-based method. As pool-based approaches are more common in perception applications, we derived a concept for comparing pool-based and stream-based AL, where TPL out-performed state-of-the-art pool- or stream-based approaches for different models. TPL demonstrated a gain of 2.5 precept points (pp) less required data while being significantly faster than pool-based methods.
Eye lens dosimetry has been an important field of research in the last decade. Dose measurements with a prototype of an active personal eye lens dosemeter based on the Dosepix detector are presented. The personal dose equivalent at $3\,$mm depth of soft tissue, $H_\text{p}(3)$, was measured in the center front of a water-filled cylinder phantom with a height and diameter of $20\,$cm. The energy dependence of the normalized response is investigated for mean photon energies between $12.4\,$keV and $248\,$keV for continuous reference radiation fields (N-series) according to ISO 4037. The response normalized to N-60 ($\overline{E}=47.9\,\text{keV}$) at $0^\circ$ angle of irradiation stays within the approval limits of IEC 61526 for angles of incidence between $-75^\circ$ and $+75^\circ$. Performance in pulsed photon fields was tested for varying dose rates from $0.1\,\frac{\text{Sv}}{\text{h}}$ up to $1000\,\frac{\text{Sv}}{\text{h}}$ and pulse durations from $1\,\text{ms}$ up to $10\,\text{s}$. The dose measurement works well within the approval limits (acc. to IEC 61526) up to $1\,\frac{\text{Sv}}{\text{h}}$. No significant influence of the pulse duration on the measured dose is found. Reproducibility measurements yield a coefficient of variation which does not exceed $1\,\%$ for two tested eye lens dosemeter prototypes.
The Archive Query Log (AQL) is a previously unused, comprehensive query log collected at the Internet Archive over the last 25 years. Its first version includes 356 million queries, 166 million search result pages, and 1.7 billion search results across 550 search providers. Although many query logs have been studied in the literature, the search providers that own them generally do not publish their logs to protect user privacy and vital business data. Of the few query logs publicly available, none combines size, scope, and diversity. The AQL is the first to do so, enabling research on new retrieval models and (diachronic) search engine analyses. Provided in a privacy-preserving manner, it promotes open research as well as more transparency and accountability in the search industry.
First investigations regarding dosimetric properties of the hybrid, pixelated, photon-counting Dosepix detector in the direct beam of a pulsed photon field (RQR8) for the personal dose equivalent $H\mathrm{_p(10)}$ are presented. The influence quantities such as pulse duration and dose rate were varied, and their responses were compared to the legal limits provided in PTB-A 23.2. The variation of pulse duration at a nearly constant dose rate of about 3.7$\,$Sv/h shows a flat response around 1.0 from 3.6$\,$s down to 2$\,$ms. A response close to 1.0 is achieved for dose rates from 0.07$\,$Sv/h to 35$\,$Sv/h for both pixel sizes. Above this dose rate, the large pixels (220$\,μ$m edge length) are below the lower limit. The small pixels (55$\,μ$m edge length) stay within limits up to 704$\,$Sv/h. The count rate linearity is compared to previous results, confirming the saturating count rate for high dose rates.
In a strongly connected graph $G = (V,E)$, a cut arc (also called strong bridge) is an arc $e \in E$ whose removal makes the graph no longer strongly connected. Equivalently, there exist $u,v \in V$, such that all $u$-$v$ walks contain $e$. Cut arcs are a fundamental graph-theoretic notion, with countless applications, especially in reachability problems. In this paper we initiate the study of cut paths, as a generalisation of cut arcs, which we naturally define as those paths $P$ for which there exist $u,v \in V$, such that all $u$-$v$ walks contain $P$ as subwalk. We first prove various properties of cut paths and define their remainder structures, which we use to present a simple $O(m)$-time verification algorithm for a cut path ($|V| = n$, $|E| = m$). Secondly, we apply cut paths and their remainder structures to improve several reachability problems from bioinformatics. A walk is called safe if it is a subwalk of every node-covering closed walk of a strongly connected graph. Multi-safety is defined analogously, by considering node-covering sets of closed walks instead. We show that cut paths provide simple $O(m)$-time algorithms verifying if a walk is safe or multi-safe. For multi-safety, we present the first linear time algorithm, while for safety, we present a simple algorithm where the state-of-the-art employed complex data structures. Finally we show that the simultaneous computation of remainder structures of all subwalks of a cut path can be performed in linear time. These properties yield an $O(mn)$ algorithm outputting all maximal multi-safe walks, improving over the state-of-the-art algorithm running in time $O(m^2+n^3)$. The results of this paper only scratch the surface in the study of cut paths, and we believe a rich structure of a graph can be revealed, considering the perspective of a path, instead of just an arc.
The diffusion of new technologies is crucial for the realization of social and economic returns to innovation. Tracking and mapping technology diffusion is, however, typically limited by the extent to which we can observe technology adoption. This study uses website texts to train a multilingual language model ensemble to map technology diffusion for the case of 3D printing. The study identifies relevant actors and their roles in the diffusion process. The results show that besides manufacturers, service provider, retailers, and information providers play an important role. The geographic distribution of adoption intensity suggests that regional 3D-printing intensity is driven by experienced lead users and the presence of technical universities. The overall adoption intensity varies by sector and firm size. These patterns indicate that the approach of using webAI provides a useful and novel tool for technology mapping which adds to existing measures based on patents or survey data.
Genome assembly is a fundamental problem in Bioinformatics, requiring to reconstruct a source genome from an assembly graph built from a set of reads (short strings sequenced from the genome). A notion of genome assembly solution is that of an arc-covering walk of the graph. Since assembly graphs admit many solutions, the goal is to find what is definitely present in all solutions, or what is safe. Most practical assemblers are based on heuristics having at their core unitigs, namely paths whose internal nodes have unit in-degree and out-degree, and which are clearly safe. The long-standing open problem of finding all the safe parts of the solutions was recently solved [RECOMB 2016] yielding a 60% increase in contig length. This safe and complete genome assembly algorithm was followed by other works improving the time bounds, as well as extending the results for different notions of assembly solution. But it remained open whether one can be complete also for models of genome assembly of practical applicability. In this paper we present a universal framework for obtaining safe and complete algorithms which unify the previous results, while also allowing for easy generalisations to assembly problems including many practical aspects. This is based on a novel graph structure, called the hydrostructure of a walk, which highlights the reachability properties of the graph from the perspective of the walk. The hydrostructure allows for simple characterisations of the existing safe walks, and of their new practical versions. Almost all of our characterisations are directly adaptable to optimal verification algorithms, and simple enumeration algorithms. Most of these algorithms are also improved to optimality using an incremental computation procedure and a previous optimal algorithm of a specific model.
In this paper we describe the methods we used for our submissions to the GermEval 2021 shared task on the identification of toxic, engaging, and fact-claiming comments. For all three subtasks we fine-tuned freely available transformer-based models from the Huggingface model hub. We evaluated the performance of various pre-trained models after fine-tuning on 80% of the training data with different hyperparameters and submitted predictions of the two best performing resulting models. We found that this approach worked best for subtask 3, for which we achieved an F1-score of 0.736.
Given a directed graph $G$ and a pair of nodes $s$ and $t$, an \emph{$s$-$t$ bridge} of $G$ is an edge whose removal breaks all $s$-$t$ paths of $G$ (and thus appears in all $s$-$t$ paths). Computing all $s$-$t$ bridges of $G$ is a basic graph problem, solvable in linear time. In this paper, we consider a natural generalisation of this problem, with the notion of "safety" from bioinformatics. We say that a walk $W$ is \emph{safe} with respect to a set $\mathcal{W}$ of $s$-$t$ walks, if $W$ is a subwalk of all walks in $\mathcal{W}$. We start by considering the maximal safe walks when $\mathcal{W}$ consists of: all $s$-$t$ paths, all $s$-$t$ trails, or all $s$-$t$ walks of $G$. We show that the first two problems are immediate linear-time generalisations of finding all $s$-$t$ bridges, while the third problem is more involved. In particular, we show that there exists a compact representation computable in linear time, that allows outputting all maximal safe walks in time linear in their length. We further generalise these problems, by assuming that safety is defined only with respect to a subset of \emph{visible} edges. Here we prove a dichotomy between the $s$-$t$ paths and $s$-$t$ trails cases, and the $s$-$t$ walks case: the former two are NP-hard, while the latter is solvable with the same complexity as when all edges are visible. We also show that the same complexity results hold for the analogous generalisations of \emph{$s$-$t$ articulation points} (nodes appearing in all $s$-$t$ paths). We thus obtain the best possible results for natural "safety"-generalisations of these two fundamental graph problems. Moreover, our algorithms are simple and do not employ any complex data structures, making them ideal for use in practice.
Given a directed graph $G$ and a pair of nodes $s$ and $t$, an $s$-$t$ bridge of $G$ is an edge whose removal breaks all $s$-$t$ paths of $G$. Similarly, an $s$-$t$ articulation point of $G$ is a node whose removal breaks all $s$-$t$ paths of $G$. Computing the sequence of all $s$-$t$ bridges of $G$ (as well as the $s$-$t$ articulation points) is a basic graph problem, solvable in linear time using the classical min-cut algorithm. When dealing with cuts of unit size ($s$-$t$ bridges) this algorithm can be simplified to a single graph traversal from $s$ to $t$ avoiding an arbitrary $s$-$t$ path, which is interrupted at the $s$-$t$ bridges. Further, the corresponding proof is also simplified making it independent of the theory of network flows.
We propose a quantum simulator based on driven superconducting qubits where the interactions are generated parametrically by a polychromatic magnetic flux modulation of a tunable bus element. Using a time-dependent Schrieffer-Wolff transformation, we analytically derive a multi-qubit Hamiltonian which features independently tunable $XX$ and $YY$-type interactions as well as local bias fields over a large parameter range. We demonstrate the adiabatic simulation of the ground state of a hydrogen molecule using two superconducting qubits and one tunable bus element. The time required to reach chemical accuracy lies in the few microsecond range and therefore could be implemented on currently available superconducting circuits. Further applications of this technique may also be found in the simulation of interacting spin systems.