Search

Noise remains one of the most significant challenges in the development of reliable and scalable quantum processors. While quantum error correction and mitigation techniques offer potential solutions, they are often limited by the substantial hardware overhead required. To address this, tailored approaches that exploit specific hardware characteristics have emerged. In quantum computing architectures utilizing cat-qubits, the inherent exponential suppression of bit-flip errors can significantly reduce the qubit count needed for effective error correction. In this work, we explore how the unique noise bias of cat-qubits can be harnessed to enhance error mitigation efficiency. Specifically, we demonstrate that the sampling cost associated with probabilistic error cancellation (PEC) methods can be substantially lowered when applied to circuits built on cat-qubits, provided the gates used preserve the noise bias. Our error mitigation scheme is benchmarked across various quantum machine learning circuits, showcasing its practical advantages., Comment: 17 pages, 11 figures, 1 algorithm

Bayesian models of cognition have gained considerable traction in computational neuroscience and psychiatry. Their scopes are now expected to expand rapidly to artificial intelligence, providing general inference frameworks to support embodied, adaptable, and energy-efficient autonomous agents. A central theory in this domain is predictive coding, which posits that learning and behaviour are driven by hierarchical probabilistic inferences about the causes of sensory inputs. Biological realism constrains these networks to rely on simple local computations in the form of precision-weighted predictions and prediction errors. This can make this framework highly efficient, but its implementation comes with unique challenges on the software development side. Embedding such models in standard neural network libraries often becomes limiting, as these libraries' compilation and differentiation backends can force a conceptual separation between optimization algorithms and the systems being optimized. This critically departs from other biological principles such as self-monitoring, self-organisation, cellular growth and functional plasticity. In this paper, we introduce \texttt{pyhgf}: a Python package backed by JAX and Rust for creating, manipulating and sampling dynamic networks for predictive coding. We improve over other frameworks by enclosing the network components as transparent, modular and malleable variables in the message-passing steps. The resulting graphs can implement arbitrary computational complexities as beliefs propagation. But the transparency of core variables can also translate into inference processes that leverage self-organisation principles, and express structure learning, meta-learning or causal discovery as the consequence of network structural adaptation to surprising inputs. The code, tutorials and documentation are hosted at: https://github.com/ilabcode/pyhgf.

To gain better understanding of the Ap stars with the longest rotation periods, we obtained high resolution spectra of a sample of super-slowly rotating Ap (ssrAp) star candidates identified by a TESS photometric survey, to confirm that they are indeed Ap stars, to check that their v sin i values are compatible with super-slow rotation, and to obtain a first estimate of their magnetic field strengths. We determined whenever possible their mean magnetic field modulus, their mean quadratic magnetic field, and an upper limit of their projected equatorial velocities. Eighteen of the 27 stars studied are typical Ap stars; most of the other nine appear to be misclassified. One of the Ap stars is not a slow rotator; it must be seen nearly pole-on. The properties of the remaining 17 are compatible with moderately to extremely long rotation periods. Eight new stars with resolved magnetically split lines in the visible range were discovered; their mean magnetic field modulus and their mean quadratic magnetic field were measured. The mean quadratic field could also be determined in five more stars. Five new spectroscopic binaries containing an Ap star were identified. Among the misclassified stars, one SB2 system with two similar, sharp-lined Am components was also discovered. The technique that we used to carry out a search for ssrAp star candidates using TESS data is validated, but appears limited by uncertainties in the spectral classification of Ap stars. The new magnetic field measurements obtained as part of this study lend further support to the tentative conclusions of our previous studies: the absence of periods longer than ~150 d in stars with magnetic fields stronger than ~7.5 kG, the lower rate of occurrence of super-slow rotation for field strengths less than ~2 kG than in the range ~3-7.5 kG, and the deficiency of slowly rotating Ap stars with field strengths between ~2 and ~3 kG., Comment: 34 pages, 30 figures, accepted for publication in A&A

Many graph algorithms can be viewed as sets of rules that are iteratively applied, with the number of iterations dependent on the size and complexity of the input graph. Existing machine learning architectures often struggle to represent these algorithmic decisions as discrete state transitions. Therefore, we propose a novel framework: GraphFSA (Graph Finite State Automaton). GraphFSA is designed to learn a finite state automaton that runs on each node of a given graph. We test GraphFSA on cellular automata problems, showcasing its abilities in a straightforward algorithmic setting. For a comprehensive empirical evaluation of our framework, we create a diverse range of synthetic problems. As our main application, we then focus on learning more elaborate graph algorithms. Our findings suggest that GraphFSA exhibits strong generalization and extrapolation abilities, presenting an alternative approach to represent these algorithms., Comment: Published as a conference paper at ECAI 2024

Artificial Intelligence (AI) tools have become incredibly powerful in generating synthetic images. Of particular concern are generated images that resemble photographs as they aspire to represent real world events. Synthetic photographs may be used maliciously by a broad range of threat actors, from scammers to nation-state actors, to deceive, defraud, and mislead people. Mitigating this threat usually involves answering a basic analytic question: Is the photograph real or synthetic? To address this, we have examined the capabilities of recent generative diffusion models and have focused on their flaws: visible artifacts in generated images which reveal their synthetic origin to the trained eye. We categorize these artifacts, provide examples, discuss the challenges in detecting them, suggest practical applications of our work, and outline future research directions., Comment: 27 pages, 25 figures

Massless spinning correlators in cosmology are extremely complicated. In contrast, the scattering amplitudes of massless particles with spin are very simple. We propose that the reason for the unreasonable complexity of these correlators lies in the use of inconvenient kinematic variables. For example, in de Sitter space, consistency with unitarity and the background isometries imply that the correlators must be conformally covariant and also conserved. However, the commonly used kinematic variables for correlators do not make all of these properties manifest. In this paper, we introduce twistor space as a powerful way to satisfy all kinematic constraints. We show that conformal correlators of conserved currents can be written as twistor integrals, where the conservation condition translates into holomorphicity of the integrand. The functional form of the twistor-space correlators is very simple and easily bootstrapped. For the case of three-point functions, we verify explicitly that this reproduces known results in embedding space. We also perform a half-Fourier transform of the twistor-space correlators to obtain their counterparts in momentum space. We conclude that twistors provide a promising new avenue to study conformal correlation functions that exposes their hidden simplicity., Comment: Typo fixed and important references added

In a four-dimensional quantum field theory that flows between two fixed points under the renormalization group, the change in the conformal anomaly $\Delta a$ has been related to the average null energy. We extend this result to derive a sum rule for the other anomaly coefficient, $\Delta c$, in terms of the stress tensor three-point function. While the sum rule for $\Delta a$ is an expectation value of the averaged null energy operator, and therefore positive, the result for $\Delta c$ involves the off-diagonal matrix elements, so it does not have a fixed sign.

The evolution of artificial intelligence (AI) has catalyzed a transformation in digital content generation, with profound implications for cyber influence operations. This report delves into the potential and limitations of generative deep learning models, such as diffusion models, in fabricating convincing synthetic images. We critically assess the accessibility, practicality, and output quality of these tools and their implications in threat scenarios of deception, influence, and subversion. Notably, the report generates content for several hypothetical cyber influence operations to demonstrate the current capabilities and limitations of these AI-driven methods for threat actors. While generative models excel at producing illustrations and non-realistic imagery, creating convincing photo-realistic content remains a significant challenge, limited by computational resources and the necessity for human-guided refinement. Our exploration underscores the delicate balance between technological advancement and its potential for misuse, prompting recommendations for ongoing research, defense mechanisms, multi-disciplinary collaboration, and policy development. These recommendations aim to leverage AI's potential for positive impact while safeguarding against its risks to the integrity of information, especially in the context of cyber influence., Comment: 44 pages, 56 figures

We explore the physics of relativistic gapless phases defined by a mixed anomaly between two generalized conserved currents. The gapless modes can be understood as Goldstone modes arising from the nonlinear realization of (generically higher-form) symmetries arising from these currents. In some cases, the anomaly cannot be reproduced by any local and unitary theory, indicating that the corresponding symmetries are impossible, in the sense that they cannot appear in a Lorentzian physical system. We consider many examples of the general construction. Most notably, we study conformal gravity from this perspective, describing the higher-form symmetries of the linear theory and showing how it can be understood in terms of anomalies. Along the way we clarify some aspects of electric-magnetic duality in linear conformal gravity., Comment: 30 pages + appendices; v2: refs added, typos corrected, published in PRD

Graph Visualization, also known as Graph Drawing, aims to find geometric embeddings of graphs that optimize certain criteria. Stress is a widely used metric; stress is minimized when every pair of nodes is positioned at their shortest path distance. However, stress optimization presents computational challenges due to its inherent complexity and is usually solved using heuristics in practice. We introduce a scalable Graph Neural Network (GNN) based Graph Drawing framework with sub-quadratic runtime that can learn to optimize stress. Inspired by classical stress optimization techniques and force-directed layout algorithms, we create a coarsening hierarchy for the input graph. Beginning at the coarsest level, we iteratively refine and un-coarsen the layout, until we generate an embedding for the original graph. To enhance information propagation within the network, we propose a novel positional rewiring technique based on intermediate node positions. Our empirical evaluation demonstrates that the framework achieves state-of-the-art performance while remaining scalable., Comment: Published as a conference paper at ICLR 2024

Pulsations of rapidly oscillating Ap stars and their interaction with the stellar magnetic field have not been studied in the near-IR region despite the benefits these observations offer compared to visual wavelengths. The main advantage of the near-IR is the quadratic dependence of the Zeeman effect on the wavelength, as opposed to the linear dependence of the Doppler effect. To test pulsation diagnostics of roAp stars in the near-IR, we aim to investigate the pulsation behaviour of one of the brightest magnetic roAp stars, $\gamma$Equ, which possesses a strong surface magnetic field of the order of several kilogauss and exhibits magnetically split spectral lines in its spectra. Two magnetically split spectral lines belonging to different elements, the triplet Fe I at 1563.63nm and the pseudo-doublet Ce III at 1629.2nm, were recorded with CRIRES+ over about one hour in the H band with the aim of understanding the character of the line profile variability and the pulsation behaviour of the magnetic field modulus. The profile shapes of both studied magnetically split spectral lines vary in a rather complex manner probably due to a significant decrease in the strength of the longitudinal field component and an increase in the strength of the transverse field components over the last decade. A mean magnetic field modulus of 3.9kG was determined for the Fe I, whereas for the Ce III we observe only about 2.9kG. For comparison, a mean field modulus of 3.4kG was determined using the Zeeman doublet Fe II at 6249.25 in optical PEPSI spectra recorded just about two weeks before the CRIRES+ observations. Different effects that may lead to the differences in the field modulus values are discussed. Our measurements of the mean magnetic field modulus using the line profiles recorded in different pulsational phase bins suggest a field modulus variability of 32G for the Fe I and 102G for the Ce III., Comment: Accepted to be published in A&A

The rotation periods of Ap stars span more than five orders of magnitude. The physical origin of this differentiation remains poorly understood. The consideration of the most slowly rotating Ap stars represents a promising approach to gain insight into the processes at play. The identification of super-slowly rotating Ap (ssrAp) star candidates (defined as Ap stars that have rotation periods longer than 50d) through systematic exploitation of the available TESS observations of Ap stars is an effective approach to build a sample devoid of magnetic bias. In our previous analyses of TESS Cycle 1 and Cycle 2 data, we interpreted the Ap stars showing no photometric variability over the 27-d duration of a TESS sector as being ssrAp star candidates. Here, we apply the same approach to TESS Cycle 3 and Cycle 4 observations of Ap stars. However, two issues may lead to spurious identification of ssrAp star candidates. (1) A considerable fraction of the Ap stars in the existing lists have erroneous or dubious spectral classifications. (2) The TESS data processing may remove part of the variability signal. After critical evaluation of these effects, we report the identification of 25 new ssrAp star candidates and of 8 stars with moderately long periods. Combining this list with the lists of ssrAp stars from Cycles 1 and 2 and with the list of ssrAp stars that were previously known but whose lack of variability was not detected in our study, we confirmed at a higher significance level the conclusions drawn in our earlier work. These include the lower rate of occurrence of super-slow rotation among weakly magnetic Ap stars than among strongly magnetic ones, the probable existence of a gap between ~2 and ~3kG in the distribution of the magnetic field strengths of the ssrAp stars, and the much higher rate of occurrence of rapid oscillations in ssrAp stars than in the whole population of Ap stars., Comment: 22 pages, 10 figures. Accepted for publication in A&A

Simulating fluid dynamics is crucial for the design and development process, ranging from simple valves to complex turbomachinery. Accurately solving the underlying physical equations is computationally expensive. Therefore, learning-based solvers that model interactions on meshes have gained interest due to their promising speed-ups. However, it is unknown to what extent these models truly understand the underlying physical principles and can generalize rather than interpolate. Generalization is a key requirement for a general-purpose fluid simulator, which should adapt to different topologies, resolutions, or thermodynamic ranges. We propose SURF, a benchmark designed to test the $\textit{generalization}$ of learned graph-based fluid simulators. SURF comprises individual datasets and provides specific performance and generalization metrics for evaluating and comparing different models. We empirically demonstrate the applicability of SURF by thoroughly investigating the two state-of-the-art graph-based models, yielding new insights into their generalization., Comment: Accepted at LoG 2023, Learning on Graphs Conference

We study applications of spectral positivity and the averaged null energy condition (ANEC) to renormalization group (RG) flows in two-dimensional quantum field theory. We find a succinct new proof of the Zamolodchikov $c$-theorem, and derive further independent constraints along the flow. In particular, we identify a natural $C$-function that is a completely monotonic function of scale, meaning its derivatives satisfy the alternating inequalities $(-1)^nC^{(n)}(\mu^2) \geq 0$. The completely monotonic $C$-function is identical to the Zamolodchikov $C$-function at the endpoints, but differs along the RG flow. In addition, we apply Lorentzian techniques that we developed recently to study anomalies and RG flows in four dimensions, and show that the Zamolodchikov $c$-theorem can be restated as a Lorentzian sum rule relating the change in the central charge to the average null energy. This establishes that the ANEC implies the $c$-theorem in two dimensions, and provides a second, simpler example of the Lorentzian sum rule., Comment: 20 pages plus appendices, 3 figures

Most Graph Neural Networks follow the standard message-passing framework where, in each step, all nodes simultaneously communicate with each other. We want to challenge this paradigm by aligning the computation more closely to the execution of distributed algorithms and propose the Flood and Echo Net. A single round of a Flood and Echo Net consists of an origin node and a flooding phase followed by an echo phase. First, during the flooding, messages are sent from the origin and propagated outwards throughout the entire graph. Then, during the echo, the message flow reverses and messages are sent back towards the origin. As nodes are only sparsely activated upon receiving a message, this leads to a wave-like activation pattern that traverses the graph. Through these sparse but parallel activations, the Net becomes more expressive than traditional MPNNs which are limited by the 1-WL test and also is provably more efficient in terms of message complexity. Moreover, the mechanism's inherent ability to generalize across graphs of varying sizes positions it as a practical architecture for the task of algorithmic learning. We test the Flood and Echo Net on a variety of synthetic tasks and the SALSA-CLRS benchmark and find that the algorithmic alignment of the execution improves generalization to larger graph sizes., Comment: 9 pages

We establish a connection between the averaged null energy condition (ANEC) and the monotonicity of the renormalization group, by studying the light-ray operator $\int du T_{uu}$ in quantum field theories that flow between two conformal fixed points. In four dimensions, we derive an exact sum rule relating this operator to the Euler coefficient in the trace anomaly, and show that the ANEC implies the a-theorem. The argument is based on matching anomalies in the stress tensor 3-point function, and relies on special properties of contact terms involving light-ray operators. We also illustrate the sum rule for the example of a free massive scalar field. Averaged null energy appears in a variety of other applications to quantum field theory, including causality constraints, Lorentzian inversion, and quantum information. The quantum information perspective provides a new derivation of the $a$-theorem from the monotonicity of relative entropy. The equation relating our sum rule to the dilaton scattering amplitude in the forward limit suggests an inversion formula for non-conformal theories., Comment: 33 pages including appendices, 3 figures

We introduce an extension to the CLRS algorithmic learning benchmark, prioritizing scalability and the utilization of sparse representations. Many algorithms in CLRS require global memory or information exchange, mirrored in its execution model, which constructs fully connected (not sparse) graphs based on the underlying problem. Despite CLRS's aim of assessing how effectively learned algorithms can generalize to larger instances, the existing execution model becomes a significant constraint due to its demanding memory requirements and runtime (hard to scale). However, many important algorithms do not demand a fully connected graph; these algorithms, primarily distributed in nature, align closely with the message-passing paradigm employed by Graph Neural Networks. Hence, we propose SALSA-CLRS, an extension of the current CLRS benchmark specifically with scalability and sparseness in mind. Our approach includes adapted algorithms from the original CLRS benchmark and introduces new problems from distributed and randomized algorithms. Moreover, we perform a thorough empirical evaluation of our benchmark. Code is publicly available at https://github.com/jkminder/SALSA-CLRS., Comment: (Extended Abstract) Presented at the Second Learning on Graphs Conference (LoG 2023)

Hierarchical Bayesian models of perception and learning feature prominently in contemporary cognitive neuroscience where, for example, they inform computational concepts of mental disorders. This includes predictive coding and hierarchical Gaussian filtering (HGF), which differ in the nature of hierarchical representations. Predictive coding assumes that higher levels in a given hierarchy influence the state (value) of lower levels. In HGF, however, higher levels determine the rate of change at lower levels. Here, we extend the space of generative models underlying HGF to include a form of nonlinear hierarchical coupling between state values akin to predictive coding and artificial neural networks in general. We derive the update equations corresponding to this generalization of HGF and conceptualize them as connecting a network of (belief) nodes where parent nodes either predict the state of child nodes or their rate of change. This enables us to (1) create modular architectures with generic computational steps in each node of the network, and (2) disclose the hierarchical message passing implied by generalized HGF models and to compare this to comparable schemes under predictive coding. We find that the algorithmic architecture instantiated by the generalized HGF is largely compatible with that of predictive coding but extends it with some unique predictions which arise from precision and volatility related computations. Our developments enable highly flexible implementations of hierarchical Bayesian models for empirical data analysis and are available as open source software.

Despite the universe containing primordial thorium (Th) of sufficient abundance to appear in stellar spectra, detection of Th has to date been tentative and based on just a few weak and blended lines. Here, we present convincing evidence not only for the first Th detection in a magnetic chemically peculiar Ap star but also for the first detection of Th III in a stellar spectrum. CPD-62 2717 was initially recognized as a highly-magnetized Ap star thanks to resolved magnetically split lines captured in $H$-band spectra from the SDSS/APOGEE survey. The star was subsequently pinpointed as extraordinarily peculiar when careful inspection of the $H$-band line content revealed the presence of five lines of Th III, none of which are detected in the other $\sim1500$ APOGEE-observed Ap stars. Follow-up with the VLT+UVES confirmed a similarly peculiar optical spectrum featuring dozens of Th III lines, among other peculiarities. Unlike past claims of Th detection, and owing to high-resolution observations of the strong ($\sim$8$-$12$\,$kG) magnetic field of CPD-62 2717, the detection of Th III can in this case be supported by matches between the observed and theoretical magnetic splitting patterns. Comparison of CPD-62 2717 to stars for which Th overabundances have been previously reported (e.g., Przybylski's Star) indicate that only for CPD-62 2717 is the Th detection certain. Along with the focus on Th III, we use time series measurements of the magnetic field modulus to constrain the rotation period of CPD-62 2717 to $\sim$4.8 years, thus establishing it as a new example of a super-slowly-rotating Ap star.

The global trends of urbanization and increased personal mobility force us to rethink the way we live and use urban space. The Traffic4cast competition series tackles this problem in a data-driven way, advancing the latest methods in machine learning for modeling complex spatial systems over time. In this edition, our dynamic road graph data combine information from road maps, $10^{12}$ probe data points, and stationary vehicle detectors in three cities over the span of two years. While stationary vehicle detectors are the most accurate way to capture traffic volume, they are only available in few locations. Traffic4cast 2022 explores models that have the ability to generalize loosely related temporal vertex data on just a few nodes to predict dynamic future traffic states on the edges of the entire road graph. In the core challenge, participants are invited to predict the likelihoods of three congestion classes derived from the speed levels in the GPS data for the entire road graph in three cities 15 min into the future. We only provide vehicle count data from spatially sparse stationary vehicle detectors in these three cities as model input for this task. The data are aggregated in 15 min time bins for one hour prior to the prediction time. For the extended challenge, participants are tasked to predict the average travel times on super-segments 15 min into the future - super-segments are longer sequences of road segments in the graph. The competition results provide an important advance in the prediction of complex city-wide traffic states just from publicly available sparse vehicle data and without the need for large amounts of real-time floating vehicle data., Comment: Pre-print under review, submitted to Proceedings of Machine Learning Research

While artificial intelligence (AI) models have achieved human or even superhuman performance in many well-defined applications, they still struggle to show signs of broad and flexible intelligence. The Abstraction and Reasoning Corpus (ARC), a visual intelligence benchmark introduced by Fran\c{c}ois Chollet, aims to assess how close AI systems are to human-like cognitive abilities. Most current approaches rely on carefully handcrafted domain-specific program searches to brute-force solutions for the tasks present in ARC. In this work, we propose a general learning-based framework for solving ARC. It is centered on transforming tasks from the vision to the language domain. This composition of language and vision allows for pre-trained models to be leveraged at each stage, enabling a shift from handcrafted priors towards the learned priors of the models. While not yet beating state-of-the-art models on ARC, we demonstrate the potential of our approach, for instance, by solving some ARC tasks that have not been solved previously., Comment: The first two authors have contributed equally to this work. Accepted as regular paper at CVPR 2023 Workshop and Challenges for New Frontiers in Visual Language Reasoning: Compositionality, Prompts and Causality (NFVLR)

We report about HD 213258, an Ap star that we recently identified as presenting a unique combination of rare, remarkable properties. Our study of this star is based on ESPaDOnS Stokes I and V data obtained at 7 epochs spanning a time interval slightly shorter than 2 years, on TESS data, and on radial velocity measurements from the CORAVEL data base. We confirm that HD 213258 is definitely an Ap star. We found that, in its spectrum, the Fe II {\lambda}6149.2 {\AA} line is resolved into its two magnetically split components. The mean magnetic field modulus of HD 213258, ~ 3.8 kG does not show significant variations over ~2 years. Comparing our mean longitudinal field determinations with a couple of measurements from the literature, we show that the stellar rotation period must likely be of the order of 50 years, with a reversal of the field polarity. Moreover, HD 213258 is a rapidly oscillating Ap (roAp) star, in which high overtone pulsations with a period of 7.58 min are detected. Finally, we confirm that HD 213258 has a mean radial velocity exceeding (in absolute value) that of at least 99% of the Ap stars. The radial velocity shows low amplitude variations, which suggests that the star is a single-line spectroscopic binary. It is also a known astrometric binary. While its orbital elements remain to be determined, its orbital period likely is one of the shortest known for a binary roAp star. Its secondary is close to the borderline between stellar and substellar objects. There is a significant probability that it may be a brown dwarf. While most of the above-mentioned properties, taken in isolation, are observed in a small fraction of the whole population of Ap stars, the probability that a single star possesses all of them is extremely low. This makes HD 213258 an exceptionally interesting object that deserves to be studied in detail in the future., Comment: Accepted for publication in A&A; 6 pages, 4 figures

Classical graph algorithms work well for combinatorial problems that can be thoroughly formalized and abstracted. Once the algorithm is derived, it generalizes to instances of any size. However, developing an algorithm that handles complex structures and interactions in the real world can be challenging. Rather than specifying the algorithm, we can try to learn it from the graph-structured data. Graph Neural Networks (GNNs) are inherently capable of working on graph structures; however, they struggle to generalize well, and learning on larger instances is challenging. In order to scale, we focus on a recurrent architecture design that can learn simple graph problems end to end on smaller graphs and then extrapolate to larger instances. As our main contribution, we identify three essential techniques for recurrent GNNs to scale. By using (i) skip connections, (ii) state regularization, and (iii) edge convolutions, we can guide GNNs toward extrapolation. This allows us to train on small graphs and apply the same model to much larger graphs during inference. Moreover, we empirically validate the extrapolation capabilities of our GNNs on algorithmic datasets., Comment: Accepted at GCLR@AAAI23, workshop on Graphs and more Complex structures for Learning and Reasoning

Traffic4cast is an annual competition to predict spatio temporal traffic based on real world data. We propose an approach using Graph Neural Networks that directly works on the road graph topology which was extracted from OpenStreetMap data. Our architecture can incorporate a hierarchical graph representation to improve the information flow between key intersections of the graph and the shortest paths connecting them. Furthermore, we investigate how the road graph can be compacted to ease the flow of information and make use of a multi-task approach to predict congestion classes and ETA simultaneously. Our code and models are released here: https://github.com/floriangroetschla/NeurIPS2022-traffic4cast