Search

The standard cosmological model currently in force, aka $\Lambda$CDM, has been plagued with a variety of tensions in the last decade or so, which puts it against the wall. At the core of the $\Lambda$CDM we have a rigid cosmological term, $\Lambda$, for the entire cosmic history. Recently, the results from the DESI collaboration suggested the possibility that dark energy (DE) should be dynamical rather than just a cosmological constant. Using a generic $w_0w_a$CDM parameterization, DESI favors quintessence behavior at $\sim 3\sigma$ c.l. However, to alleviate the tensions the DE needs more features. In the proposed $w$XCDM model of [45] the DE is actually a composite cosmic fluid with two components $(X,Y)$ acting sequentially: first $X$ (above a transition redshift $z_t$) and second $Y$ (below $z_t$). Fitting the model to the data, we find that the late component $Y$ behaves as quintessence, like DESI. However, to cure the $H_0$ and growth tensions, $X$ must behave as `phantom matter' (PM), which in contrast to phantom DE provides positive pressure at the expense of negative energy density. Using the SNIa (considering separately Pantheon$+$ and DESY5), cosmic chronometers, transversal BAO, LSS data, and the full CMB likelihood from Planck 2018, we find that both tensions can be completely cut down. We also compare the $w$XCDM with our own results using the standard $w$CDM and $w_0w_a$CDM parameterizations of the DE. In all cases, model $w$XCDM performs much better. Finally, we have repeated our analysis with BAO 3D data (replacing BAO 2D), and we still find that the main dynamical DE models (including composite ones) provide a much better fit quality compared to $\Lambda$CDM. The growth tension is alleviated again, but in contrast, the $H_0$-tension remains significant, which is most likely reminiscent of the internal conflict in the BAO sector., Comment: 25 pages, 3 Figures and 2 Tables. Typos corrected, references added

This paper proposes a data-driven algorithm for model order reduction (MOR) of large-scale wind farms and studies the effects that the obtained reduced-order model (ROM) has when this is integrated into the power grid. With respect to standard MOR methods, the proposed algorithm has the advantages of having low computational complexity and not requiring any knowledge of the high order model. Using time-domain measurements, the obtained ROM achieves the moment matching conditions at selected interpolation points (frequencies). With respect to the state of the art, the method achieves the so-called two-sided moment matching, doubling the accuracy by doubling the interpolated points. The proposed algorithm is validated on a combined model of a 200-turbine wind farm (which is reduced) interconnected to the IEEE 14-bus system (which represents the unreduced study area) by comparing the full-order model and the reduced-order model in terms of their Bode plots, eigenvalues and the point of common coupling voltages in extensive fault scenarios of the integrated power system., Comment: This article has been accepted for publication by IEEE Transactions on Power Systems

Pair creation is a fundamental prediction of quantum field theory in curved spacetimes. While classical aspects of this phenomenon have been observed, the experimental confirmation of its quantum origin remains elusive. In this article, we quantify the entanglement produced by pair creation in a two dimensional Bose-Einstein Condensate (BEC) analogues of expanding universes and examine the impact of various experimental factors, including decoherence from thermal noise and losses. Our analysis evaluates the feasibility of detecting entanglement in these systems and identifies optimal experimental configurations for achieving this goal. Focusing on the experimental setup detailed in \cite{Viermann:2022wgw}, we demonstrate that entanglement can be observed in these BEC analogues at a significance level of $\sim 2\sigma$ with current capabilities, and at $\gtrsim 3.3\sigma$ with minor improvements. Achieving this would provide unequivocal evidence of the quantum nature of pair creation and validate one of the most iconic predictions of quantum field theory in curved spacetimes., Comment: 17 pages, 10 Figures, comments are welcome and encouraged

We present a general framework for incorporating non-reciprocal interactions into the Ising model with Glauber dynamics, without requiring multiple species. We then focus on a model with vision-cone type interactions. We solve it in a fully connected network (mean-field) and perform extensive numerical simulations of the model in the square lattice. We find that the breakdown of the spin-flip symmetry introduced by non-reciprocity induces a discontinuous phase transition on top of the usual continuous one, that eventually occurs at higher critical temperatures. Combining a static and dynamic scaling analysis, we measure the critical exponents associated to the continuous symmetry breaking transition, and find them to be identical to the ones of the Ising model in two dimensions (2D), with the exception of the exponent $\beta$ associated to the order parameter. The latter appears to increase as the non-reciprocity of the coupling increases, suggesting that, within our numerical precision, the model does not belong to the 2D Ising model universality class. The coarsening process is anisotropic, but still follows the usual dynamic scaling with an exponent compatible with the standard value with non-conserved order parameter dynamics., Comment: 23 pages, 10 figures

Historical ciphered manuscripts are documents that were typically used in sensitive communications within military and diplomatic contexts or among members of secret societies. These secret messages were concealed by inventing a method of writing employing symbols from diverse sources such as digits, alchemy signs and Latin or Greek characters. When studying a new, unseen cipher, the automatic search and grouping of ciphers with a similar alphabet can aid the scholar in its transcription and cryptanalysis because it indicates a probability that the underlying cipher is similar. In this study, we address this need by proposing the CSI metric, a novel way of comparing pairs of ciphered documents. We assess their effectiveness in an unsupervised clustering scenario utilising visual features, including SIFT, pre-trained learnt embeddings, and OCR descriptors., Comment: Acccepted at ECCV24 Workshop AI4DH

Quantum optical networks are instrumental to address fundamental questions and enable applications ranging from communication to computation and, more recently, machine learning. In particular, photonic artificial neural networks offer the opportunity to exploit the advantages of both classical and quantum optics. Photonic neuro-inspired computation and machine learning have been successfully demonstrated in classical settings, while quantum optical networks have triggered breakthrough applications such as teleportation, quantum key distribution and quantum computing. We present a perspective on the state of the art in quantum optical machine learning and the potential advantages of artificial neural networks in circuit designs and beyond, in more general analogue settings characterised by recurrent and coherent complex interactions. We consider two analogue neuro-inspired applications, namely quantum reservoir computing and quantum associative memories, and discuss the enhanced capabilities offered by quantum substrates, highlighting the specific role of light squeezing in this context.

During clinical treatment for epilepsy, the area of the brain thought to be responsible for pathological activity is identified. This identification is typically performed through visual assessment of EEG recordings; however, this is time consuming and prone to subjective inconsistency. Automated onset localisation algorithms provide objective identification of the onset location by highlighting changes in signal features associated with seizure onset. In this work we investigate how methodological differences in such algorithms can result in different onset locations being identified. We analysed ictal intracranial EEG (icEEG) recordings in 16 subjects (100 seizures) with drug-resistant epilepsy from the SWEZ-ETHZ public database. We identified a series of key methodological differences that must be considered when designing or selecting an onset localisation algorithm. These differences were demonstrated using three distinct algorithms that capture different, but complementary, seizure onset features: Imprint, Epileptogenicity Index, and Low Entropy Map. We assessed methodological differences (or Decision Points), and their impact on the identified onset locations. Our independent application of all three algorithms to the same ictal icEEG dataset revealed low agreement between them: 27-60% of onset channels showed minimal or no overlap. Therefore, we investigated the effect of three key differences: (i) how to define a baseline, (ii) whether low-frequency components are considered, and finally (iii) whether electrodecrement is considered. Changes at each Decision Point were found to substantially influence resultant onset channels (r>0.3). Our results demonstrate how seemingly small methodological changes can result in large differences in onset locations. We propose that key Decision Points must be considered when using or designing an onset localisation algorithm.

Large language models (LLMs) demonstrate surprising capabilities, but we do not understand how they are implemented. One hypothesis suggests that these capabilities are primarily executed by small subnetworks within the LLM, known as circuits. But how can we evaluate this hypothesis? In this paper, we formalize a set of criteria that a circuit is hypothesized to meet and develop a suite of hypothesis tests to evaluate how well circuits satisfy them. The criteria focus on the extent to which the LLM's behavior is preserved, the degree of localization of this behavior, and whether the circuit is minimal. We apply these tests to six circuits described in the research literature. We find that synthetic circuits -- circuits that are hard-coded in the model -- align with the idealized properties. Circuits discovered in Transformer models satisfy the criteria to varying degrees. To facilitate future empirical studies of circuits, we created the \textit{circuitry} package, a wrapper around the \textit{TransformerLens} library, which abstracts away lower-level manipulations of hooks and activations. The software is available at \url{https://github.com/blei-lab/circuitry}., Comment: Code available here: https://github.com/blei-lab/circuitry

Recent advances in differentially private federated learning (DPFL) algorithms have found that using correlated noise across the rounds of federated learning (DP-FTRL) yields provably and empirically better accuracy than using independent noise (DP-SGD). While DP-SGD is well-suited to federated learning with a single untrusted central server using lightweight secure aggregation protocols, secure aggregation is not conducive to implementing modern DP-FTRL techniques without assuming a trusted central server. DP-FTRL based approaches have already seen widespread deployment in industry, albeit with a trusted central curator who provides and applies the correlated noise. To realize a fully private, single untrusted server DP-FTRL federated learning protocol, we introduce secure stateful aggregation: a simple append-only data structure that allows for the private storage of aggregate values and reading linear functions of the aggregates. Assuming Ring Learning with Errors, we provide a lightweight and scalable realization of this protocol for high-dimensional data in a new security/resource model, Federated MPC : where a powerful persistent server interacts with weak, ephemeral clients. We observe that secure stateful aggregation suffices for realizing DP-FTRL-based private federated learning: improving DPFL utility guarantees over the state of the art while maintaining privacy with an untrusted central party. Our approach has minimal overhead relative to existing techniques which do not yield comparable utility. The secure stateful aggregation primitive and the federated MPC paradigm may be of interest for other practical applications.

Long-form question answering (LFQA) aims at generating in-depth answers to end-user questions, providing relevant information beyond the direct answer. However, existing retrievers are typically optimized towards information that directly targets the question, missing out on such contextual information. Furthermore, there is a lack of training data for relevant context. To this end, we propose and compare different weak supervision techniques to optimize retrieval for contextual information. Experiments demonstrate improvements on the end-to-end QA performance on ASQA, a dataset for long-form question answering. Importantly, as more contextual information is retrieved, we improve the relevant page recall for LFQA by 14.7% and the groundedness of generated long-form answers by 12.5%. Finally, we show that long-form answers often anticipate likely follow-up questions, via experiments on a conversational QA dataset., Comment: Accepted at EMNLP 2024 (Findings)

Particle-producing targets in high-energy research facilities are often made from refractory metals, and they typically require dedicated cooling systems due to the challenging thermomechanical conditions they experience. However, direct contact of water with target blocks can induce erosion, corrosion, and embrittlement, especially of tungsten (W). One approach to overcoming this problem is cladding the blocks with tantalum (Ta). Unfortunately, Ta generates high decay heat when irradiated, raising safety concerns in the event of a loss-of-cooling accident. This study explored the capacity of niobium (Nb) and its alloys to form diffusion bonds with W and TZM (a molybdenum alloy with titanium and zirconium). This is because the Beam Dump Facility (BDF), a planned new fixed-target installation in CERN's North Area, uses these target materials. The bonding quality of pure Nb, Nb1Zr, and C103 (a Nb alloy with 10% hafnium and 1% titanium) with TZM and W obtained using hot isostatic pressing (HIP) was evaluated. The effects of different HIP temperatures and the introduction of a Ta interlayer were examined. Optical microscopy indicated promising bonding interfaces, which were further characterized using tensile tests and thermal-diffusivity measurements. Their performance under high-energy beam impact was validated using thermomechanical simulations. C103 exhibited higher interface strengths and safety factors than Ta2.5W, positioning it as a potential alternative cladding material for the BDF production target. The findings highlight the viability of Nb-based materials, particularly C103, for improving operational safety and efficiency in fixed-target physics experiments; however, considerations regarding the long half-life of 94Nb require further attention.

The Beam Dump Facility (BDF) is a planned fixed-target installation in CERN's North Area, set to start operating in 2031. A proton beam of 400 GeV/c will be delivered in 1 s pulses of 4e13 protons every 7.2 s, amounting to 4e19 protons on target (PoT) per year, with an average thermal power deposition of 305 kW. The experiment requires high-Z and high-density materials and involves challenging thermomechanical conditions; hence, a water-cooled refractory metal core is proposed. To prevent corrosion and erosion, the target materials - tungsten and a molybdenum-based alloy, TZM-, are clad with a tantalum alloy using hot isostatic pressing. To verify the reliability of the target design and manufacturing process, a reduced-scale prototype target was manufactured and irradiated with 2.4e16 PoT in 2018. This paper presents the findings of the post-irradiation examination (PIE) conducted 2.5 years later. The PIE consisted of non-destructive and destructive activities, including film imaging, microscopy, metrology, ultrasonic testing, microstructural analysis, and mechanical and thermal characterization. No irradiation effects resulting from thermal stresses, such as geometrical changes or cracks in the cladding, were observed, and it was determined that the bonding between the core and cladding materials was robust. Moreover, no changes were detected in the microstructures or mechanical properties of the bulk materials. In conclusion, the robustness of the target baseline design and manufacturing process was confirmed, validating its suitability for operating under the desired conditions. Nonetheless, the observed brittleness of sintered tungsten indicates a potential area for improvement to enhance the lifetime of the BDF target. Overall, the examinations provided valuable insights into the performance of the prototype and indicated potential refinements for the future BDF target complex.

One of the most captivating properties of polaritons is their capacity to confine light at the nanoscale. This confinement is even more extreme in two-dimensional (2D) materials. 2D polaritons have been investigated by optical measurements using an external photodetector. However, their effective spectrally resolved electrical detection via far-field excitation remains unexplored. This fact hinders their potential exploitation in crucial applications such as sensing molecules and gases, hyperspectral imaging and optical spectrometry, banking on their potential for integration with silicon technologies. Herein, we present the first electrical spectroscopy of polaritonic nanoresonators based on a high-quality 2D-material heterostructure, which serves at the same time as the photodetector and the polaritonic platform. We employ metallic nanorods to create hybrid nanoresonators within the hybrid plasmon-phonon polaritonic medium in the mid and long-wave infrared ranges. Subsequently, we electrically detect these resonators by near-field coupling to a graphene pn-junction. The nanoresonators simultaneously present a record of lateral confinement and high-quality factors of up to 200, exhibiting prominent peaks in the photocurrent spectrum, particularly at the underexplored lower reststrahlen band of hBN. We exploit the geometrical and gate tunability of these nanoresonators to investigate their impact on the photocurrent spectrum and the polaritonic's waveguided modes. This work opens a venue for studying this highly tunable and complex hybrid system, as well as for using it in compact platforms for sensing and photodetection applications., Comment: 34 pages, 4 main figures and 22 supplementary figures

Identifying faulty lines and their accurate location is key for rapidly restoring distribution systems. This will become a greater challenge as the penetration of power electronics increases, and contingencies are seen across larger areas. This paper proposes a single terminal methodology (i.e., no communication involved) that is robust to variations of key parameters (e.g., sampling frequency, system parameters, etc.) and performs particularly well for low resistance faults that constitute the majority of faults in low voltage DC systems. The proposed method uses local measurements to estimate the current caused by the other terminals affected by the contingency. This mimics the strategy followed by double terminal methods that require communications and decouples the accuracy of the methodology from the fault resistance. The algorithm takes consecutive voltage and current samples, including the estimated current of the other terminal, into the analysis. This mathematical methodology results in a better accuracy than other single-terminal approaches found in the literature. The robustness of the proposed strategy against different fault resistances and locations is demonstrated using MATLAB simulations., Comment: SEST 2024

The dawn of Foundation Models has on the one hand revolutionised a wide range of research problems, and, on the other hand, democratised the access and use of AI-based tools by the general public. We even observe an incursion of these models into disciplines related to human psychology, such as the Affective Computing domain, suggesting their affective, emerging capabilities. In this work, we aim to raise awareness of the power of Foundation Models in the field of Affective Computing by synthetically generating and analysing multimodal affective data, focusing on vision, linguistics, and speech (acoustics). We also discuss some fundamental problems, such as ethical issues and regulatory aspects, related to the use of Foundation Models in this research area.

The use of image analysis in automated photography management is an increasing trend in heritage institutions. Such tools alleviate the human cost associated with the manual and expensive annotation of new data sources while facilitating fast access to the citizenship through online indexes and search engines. However, available tagging and description tools are usually designed around modern photographs in English, neglecting historical corpora in minoritized languages, each of which exhibits intrinsic particularities. The primary objective of this research is to study the quantitative contribution of generative systems in the description of historical sources. This is done by contextualizing the task of captioning historical photographs from the Catalan archives as a case study. Our findings provide practitioners with tools and directions on transfer learning for captioning models based on visual adaptation and linguistic proximity., Comment: Accepted at ECCV workshop AI4DH

Current cosmological observations allow for deviations from the standard growth of large-scale structures in the universe. These deviations could indicate modifications to General Relativity on cosmological scales or suggest the dynamical nature of dark energy. It is important to characterize these departures in a model-independent manner to understand their significance objectively and explore their fundamental causes more generically across a wider spectrum of theories and models. In this paper, we compress the information from redshift-space distortion data into 2-3 parameters $\mu_i$, which control the ratio between the effective gravitational coupling in Poisson's equation and Newton's constant in several redshift bins in the late universe. We test the efficiency of this compression using mock final-year data from the Dark Energy Spectroscopic Instrument (DESI) and considering three different models within the class of effective field theories of dark energy. The constraints on the parameters of these models, obtained from both the direct fit to the data and the projection of the compressed parameters onto the parameters of the models, are fully consistent, demonstrating the method's good performance. Then, we apply it to current data and find hints of a suppressed matter growth in the universe at $\sim 2.7\sigma$ C.L., in full accordance with previous works in the literature. Finally, we perform a forecast with DESI data and show that the uncertainties on the parameters $\mu_1$ at $z<1$ and $\mu_2$ at $1

Associative memory refers to the ability to relate a memory with an input and targets the restoration of corrupted patterns. It has been intensively studied in classical physical systems, as in neural networks where an attractor dynamics settles on stable solutions. Several extensions to the quantum domain have been recently reported, displaying different features. In this work, we develop a comprehensive framework for a quantum associative memory based on open quantum system dynamics, which allows us to compare existing models, identify the theoretical prerequisites for performing associative memory tasks, and extend it in different forms. The map that achieves an exponential increase in the number of stored patterns with respect to classical systems is derived. We establish the crucial role of symmetries and dissipation in the operation of quantum associative memory. Our theoretical analysis demonstrates the feasibility of addressing both quantum and classical patterns, orthogonal and non-orthogonal memories, stationary and metastable operating regimes, and measurement-based outputs. Finally, this opens up new avenues for practical applications in quantum computing and machine learning, such as quantum error correction or quantum memories.

Multivariate Time Series Classification (MTSC) is a ubiquitous problem in science and engineering, particularly in neuroscience, where most data acquisition modalities involve the simultaneous time-dependent recording of brain activity in multiple brain regions. In recent years, Random Convolutional Kernel models such as ROCKET and MiniRocket have emerged as highly effective time series classification algorithms, capable of achieving state-of-the-art accuracy results with low computational load. Despite their success, these types of models face two major challenges when employed in neuroscience: 1) they struggle to deal with high-dimensional data such as EEG and MEG, and 2) they are difficult to interpret. In this work, we present a novel ROCKET-based algorithm, named Detach-Rocket Ensemble, that is specifically designed to address these two problems in MTSC. Our algorithm leverages pruning to provide an integrated estimation of channel importance, and ensembles to achieve better accuracy and provide a label probability. Using a synthetic multivariate time series classification dataset in which we control the amount of information carried by each of the channels, we first show that our algorithm is able to correctly recover the channel importance for classification. Then, using two real-world datasets, a MEG dataset and an EEG dataset, we show that Detach-Rocket Ensemble is able to provide both interpretable channel relevance and competitive classification accuracy, even when applied directly to the raw brain data, without the need for feature engineering., Comment: To be published in European Conference on Machine Learning and Data Mining 2024, 20 pages, 7 figures, 2 tables

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

In recent years, deep learning methods have shown impressive results for camera-based remote physiological signal estimation, clearly surpassing traditional methods. However, the performance and generalization ability of Deep Neural Networks heavily depends on rich training data truly representing different factors of variation encountered in real applications. Unfortunately, many current remote photoplethysmography (rPPG) datasets lack diversity, particularly in darker skin tones, leading to biased performance of existing rPPG approaches. To mitigate this bias, we introduce PhysFlow, a novel method for augmenting skin diversity in remote heart rate estimation using conditional normalizing flows. PhysFlow adopts end-to-end training optimization, enabling simultaneous training of supervised rPPG approaches on both original and generated data. Additionally, we condition our model using CIELAB color space skin features directly extracted from the facial videos without the need for skin-tone labels. We validate PhysFlow on publicly available datasets, UCLA-rPPG and MMPD, demonstrating reduced heart rate error, particularly in dark skin tones. Furthermore, we demonstrate its versatility and adaptability across different data-driven rPPG methods.

Nanoparticle surface structural dynamics is believed to play a significant role in regulating functionalities such as diffusion, reactivity, and catalysis but the atomic-level processes are not well understood. Atomic resolution characterization of nanoparticle surface dynamics is challenging since it requires both high spatial and temporal resolution. Though ultrafast transmission electron microscopy (TEM) can achieve picosecond temporal resolution, it is limited to nanometer spatial resolution. On the other hand, with the high readout rate of new electron detectors, conventional TEM has the potential to visualize atomic structure with millisecond time resolutions. However, the need to limit electron dose rates to reduce beam damage yields millisecond images that are dominated by noise, obscuring structural details. Here we show that a newly developed unsupervised denoising framework based on artificial intelligence enables observations of metal nanoparticle surfaces with time resolutions down to 10 ms at moderate electron dose. On this timescale, we find that many nanoparticle surfaces continuously transition between ordered and disordered configurations. The associated stress fields can penetrate below the surface leading to defect formation and destabilization making the entire nanoparticle fluxional. Combining this unsupervised denoiser with electron microscopy greatly improves spatio-temporal characterization capabilities, opening a new window for future exploration of atomic-level structural dynamics in materials.

How a neural network (NN) generalizes to novel situations depends on whether it has learned to select actions heuristically or via a planning process. "An investigation of model-free planning" (Guez et al. 2019) found that a recurrent NN (RNN) trained to play Sokoban appears to plan, with extra computation steps improving the RNN's success rate. We replicate and expand on their behavioral analysis, finding the RNN learns to give itself extra computation steps in complex situations by "pacing" in cycles. Moreover, we train linear probes that predict the future actions taken by the network and find that intervening on the hidden state using these probes controls the agent's subsequent actions. Leveraging these insights, we perform model surgery, enabling the convolutional NN to generalize beyond its 10x10 architectural limit to arbitrarily sized inputs. The resulting model solves challenging, highly off-distribution levels. We open-source our model and code, and believe the neural network's small size (1.29M parameters) makes it an excellent model organism to deepen our understanding of learned planning., Comment: Mechanistic Interpretability workshop, ICML 2024

Circuits are supposed to accurately describe how a neural network performs a specific task, but do they really? We evaluate three circuits found in the literature (IOI, greater-than, and docstring) in an adversarial manner, considering inputs where the circuit's behavior maximally diverges from the full model. Concretely, we measure the KL divergence between the full model's output and the circuit's output, calculated through resample ablation, and we analyze the worst-performing inputs. Our results show that the circuits for the IOI and docstring tasks fail to behave similarly to the full model even on completely benign inputs from the original task, indicating that more robust circuits are needed for safety-critical applications., Comment: 19 pages, 10 figures

When applying reinforcement learning from human feedback (RLHF), the reward is learned from data and, therefore, always has some error. It is common to mitigate this by regularizing the policy with KL divergence from a base model, with the hope that balancing reward with regularization will achieve desirable outcomes despite this reward misspecification. We show that when the reward function has light-tailed error, optimal policies under less restrictive KL penalties achieve arbitrarily high utility. However, if error is heavy-tailed, some policies obtain arbitrarily high reward despite achieving no more utility than the base model--a phenomenon we call catastrophic Goodhart. We adapt a discrete optimization method to measure the tails of reward models, finding that they are consistent with light-tailed error. However, the pervasiveness of heavy-tailed distributions in many real-world applications indicates that future sources of RL reward could have heavy-tailed error, increasing the likelihood of reward hacking even with KL regularization., Comment: Mechanistic Interpretability workshop at ICML 2024; Main conference poster at NeurIPS 2024

Mechanistic interpretability methods aim to identify the algorithm a neural network implements, but it is difficult to validate such methods when the true algorithm is unknown. This work presents InterpBench, a collection of semi-synthetic yet realistic transformers with known circuits for evaluating these techniques. We train simple neural networks using a stricter version of Interchange Intervention Training (IIT) which we call Strict IIT (SIIT). Like the original, SIIT trains neural networks by aligning their internal computation with a desired high-level causal model, but it also prevents non-circuit nodes from affecting the model's output. We evaluate SIIT on sparse transformers produced by the Tracr tool and find that SIIT models maintain Tracr's original circuit while being more realistic. SIIT can also train transformers with larger circuits, like Indirect Object Identification (IOI). Finally, we use our benchmark to evaluate existing circuit discovery techniques.

How well will current interpretability techniques generalize to future models? A relevant case study is Mamba, a recent recurrent architecture with scaling comparable to Transformers. We adapt pre-Mamba techniques to Mamba and partially reverse-engineer the circuit responsible for the Indirect Object Identification (IOI) task. Our techniques provide evidence that 1) Layer 39 is a key bottleneck, 2) Convolutions in layer 39 shift names one position forward, and 3) The name entities are stored linearly in Layer 39's SSM. Finally, we adapt an automatic circuit discovery tool, positional Edge Attribution Patching, to identify a Mamba IOI circuit. Our contributions provide initial evidence that circuit-based mechanistic interpretability tools work well for the Mamba architecture.

Steering vectors (SVs) have been proposed as an effective approach to adjust language model behaviour at inference time by intervening on intermediate model activations. They have shown promise in terms of improving both capabilities and model alignment. However, the reliability and generalisation properties of this approach are unknown. In this work, we rigorously investigate these properties, and show that steering vectors have substantial limitations both in- and out-of-distribution. In-distribution, steerability is highly variable across different inputs. Depending on the concept, spurious biases can substantially contribute to how effective steering is for each input, presenting a challenge for the widespread use of steering vectors. Out-of-distribution, while steering vectors often generalise well, for several concepts they are brittle to reasonable changes in the prompt, resulting in them failing to generalise well. Overall, our findings show that while steering can work well in the right circumstances, there remain technical difficulties of applying steering vectors to guide models' behaviour at scale. Our code is available at https://github.com/dtch1997/steering-bench

The growth of quantum technologies is attracting the interest of many students eager to learn concepts such as quantum entanglement or quantum superposition. However, the non-intuitive nature of these concepts poses a challenge to understanding them. Here, we present an entangled photon system which can perform a Bell test, i.e. the CHSH inequality, and can obtain the complete tomography of the two-photon state. The proposed setup is versatile, cost-effective and allows for multiple classroom operating modes. We present two variants, both facilitating the measurement of Bell inequalities and quantum state tomography. Experimental results showcase successful manipulation of the quantum state of the photons, achieving high-fidelity entangled states and significant violations of Bell's inequalities. Our setup's simplicity and affordability enhances accessibility for less specialized laboratories, allowing students to familiarize themselves with quantum physics concepts.

This research dissects financial equity research reports (ERRs) by mapping their content into categories. There is insufficient empirical analysis of the questions answered in ERRs. In particular, it is not understood how frequently certain information appears, what information is considered essential, and what information requires human judgment to distill into an ERR. The study analyzes 72 ERRs sentence-by-sentence, classifying their 4940 sentences into 169 unique question archetypes. We did not predefine the questions but derived them solely from the statements in the ERRs. This approach provides an unbiased view of the content of the observed ERRs. Subsequently, we used public corporate reports to classify the questions' potential for automation. Answers were labeled "text-extractable" if the answers to the question were accessible in corporate reports. 78.7% of the questions in ERRs can be automated. Those automatable question consist of 48.2% text-extractable (suited to processing by large language models, LLMs) and 30.5% database-extractable questions. Only 21.3% of questions require human judgment to answer. We empirically validate using Llama-3-70B and GPT-4-turbo-2024-04-09 that recent advances in language generation and information extraction enable the automation of approximately 80% of the statements in ERRs. Surprisingly, the models complement each other's strengths and weaknesses well. The research confirms that the current writing process of ERRs can likely benefit from additional automation, improving quality and efficiency. The research thus allows us to quantify the potential impacts of introducing large language models in the ERR writing process. The full question list, including the archetypes and their frequency, will be made available online after peer review., Comment: JEL classes: C45; G11; G12; G14

Traditional nanophotonic sensing schemes utilize evanescent fields in dielectric or metallic nanoparticles, which confine far-field radiation in dispersive and lossy media. Apart from the lack of a well-defined sensing volume that can be accompanied by moderate sensitivities, these structures suffer from the generally limited access to the modal field, which is key for sensing performance. Recently, a novel strategy for dielectric nanophotonics has been demonstrated, namely, the resonant confinement of light in air. So-called Mie voids created in high-index dielectric host materials support localized resonant modes with exceptional properties. In particular, due to the confinement in air, these structures benefit from the full access to the modal field inside the void. We utilize these Mie voids for refractive index sensing in single voids with volumes down to 100 attoliters and sensitivities on the order of 400 nm per refractive index unit. Taking the signal-to-noise ratio of our measurements into account, we demonstrate detection of refractive index changes as small as 6.9 x 10-4 in a defined volume of just 850 attoliters. The combination of our Mie void sensor platform with appropriate surface functionalization will even enable specificity to biological or other analytes of interest, as the sensing volumes are on the order of cellular signaling chemicals of single vesicles in cellular synapses.

This paper introduces Fetch-A-Set (FAS), a comprehensive benchmark tailored for legislative historical document analysis systems, addressing the challenges of large-scale document retrieval in historical contexts. The benchmark comprises a vast repository of documents dating back to the XVII century, serving both as a training resource and an evaluation benchmark for retrieval systems. It fills a critical gap in the literature by focusing on complex extractive tasks within the domain of cultural heritage. The proposed benchmark tackles the multifaceted problem of historical document analysis, including text-to-image retrieval for queries and image-to-text topic extraction from document fragments, all while accommodating varying levels of document legibility. This benchmark aims to spur advancements in the field by providing baselines and data for the development and evaluation of robust historical document retrieval systems, particularly in scenarios characterized by wide historical spectrum., Comment: Preprint for the manuscript accepted for publication in the DAS2024 LNCS proceedings

Looking at the well-known Hubble tension as a tension in the calibrators of the cosmic distance ladder, i.e. the absolute magnitude $M$ of standard candles such as supernovae of Type Ia (SNIa) and the standard ruler represented by the comoving sound horizon at the baryon-drag epoch, $r_d$, we propose a model-independent method to measure these distance calibrators independently from the cosmic microwave background and the first rungs of the direct distance ladder. To do so, we leverage state-of-the-art data on cosmic chronometers (CCH), SNIa and baryon acoustic oscillations (BAO) from various galaxy surveys. Taking advantage of the Gaussian Processes Bayesian technique, we reconstruct $M(z)$, $\Omega_k(z)$ and $r_d(z)$ at $z\lesssim2$ and check that no significant statistical evolution is preferred at 68\% C.L. This allows us to treat them as constants and constrain them assuming the metric description of gravity, the cosmological principle and the validity of CCH as reliable cosmic clocks, and SNIa and BAO as optimal standard candles and standard rulers, respectively, but otherwise in a model-independent way. We obtain: $\Omega_k=-0.07^{+0.12}_{-0.15}$, $M=(-19.314^{+0.086}_{-0.108})$ mag and $r_d=(142.3\pm 5.3)$ Mpc. At present, the uncertainties derived are still too large to arbitrate the tension but this is bound to change in the near future with the advent of upcoming surveys and data., Comment: 4 pages, 3 figures, 1 table. Contribution to the 2024 Cosmology session of the 58th Rencontres de Moriond

Several studies in the literature have found a disagreement between data on Baryon Acoustic Oscillations (BAO) derived using two distinct methodologies: the two-dimensional (2D or angular) BAO, which extracts the BAO signal from the angular two-point correlation function; and the three-dimensional (3D) BAO, which also exploits the radial signal imprinted on the large-scale structure of the universe. This discrepancy is worrisome, since many of the points contained in these data sets are obtained from the same catalogs of tracers, so we would expect them to be consistent. Since BAO measurements play a pivotal role in the building of the inverse distance ladder, this mismatch impacts the discourse on the Hubble tension and the theoretical solutions to the latter. So far, the discrepancy between 2D and 3D BAO has been only pointed out in the context of fitting analyses of cosmological models or parametrizations that involve a concrete calibration of the comoving sound horizon at the baryon-drag epoch. In this Letter, we avoid the use of any calibration and cosmological model, assuming that the Etherington (a.k.a distance duality) relation holds. We use state-of-the-art measurements in our analysis, and study how the results change when the angular components of the 3D BAO data from BOSS/eBOSS are substituted by the recent data from DESI Y1. We find the tension to exist at the level of $\sim 2\sigma$ and $\sim 2.5\sigma$, respectively, when the SNIa of Pantheon+ are used, and at $\sim 4.6\sigma$ when they are replaced with those of DES Y5. We then apply a calibrator-independent method to investigate the robustness of the distance duality relation when analyzed not only with 3D BAO measurements, but also with 2D BAO. We do not find any hint for a violation of the cosmic distance duality relation in any of the considered data sets. [abridged], Comment: 11 pages, 3 figures, 3 tables. Version accepted for publication in Phys. Lett. B

The Memory stress (Mess) framework provides a unified view of the memory system benchmarking, simulation and application profiling. The Mess benchmark provides a holistic and detailed memory system characterization. It is based on hundreds of measurements that are represented as a family of bandwidth--latency curves. The benchmark increases the coverage of all the previous tools and leads to new findings in the behavior of the actual and simulated memory systems. We deploy the Mess benchmark to characterize Intel, AMD, IBM, Fujitsu, Amazon and NVIDIA servers with DDR4, DDR5, HBM2 and HBM2E memory. The Mess memory simulator uses bandwidth--latency concept for the memory performance simulation. We integrate Mess with widely-used CPUs simulators enabling modeling of all high-end memory technologies. The Mess simulator is fast, easy to integrate and it closely matches the actual system performance. By design, it enables a quick adoption of new memory technologies in hardware simulators. Finally, the Mess application profiling positions the application in the bandwidth--latency space of the target memory system. This information can be correlated with other application runtime activities and the source code, leading to a better overall understanding of the application's behavior. The current Mess benchmark release covers all major CPU and GPU ISAs, x86, ARM, Power, RISC-V, and NVIDIA's PTX. We also release as open source the ZSim, gem5 and OpenPiton Metro-MPI integrated with the Mess simulator for DDR4, DDR5, Optane, HBM2, HBM2E and CXL memory expanders. The Mess application profiling is already integrated into a suite of production HPC performance analysis tools., Comment: 20 pages; accepted in MICRO-57 (2024), Best paper runner-up award

Recent advancements in data-driven approaches for remote photoplethysmography (rPPG) have significantly improved the accuracy of remote heart rate estimation. However, the performance of such approaches worsens considerably under video compression, which is nevertheless necessary to store and transmit video data efficiently. In this paper, we present a novel approach to address the impact of video compression on rPPG estimation, which leverages a pulse-signal magnification transformation to adapt compressed videos to an uncompressed data domain in which the rPPG signal is magnified. We validate the effectiveness of our model by exhaustive evaluations on two publicly available datasets, UCLA-rPPG and UBFC-rPPG, employing both intra- and cross-database performance at several compression rates. Additionally, we assess the robustness of our approach on two additional highly compressed and widely-used datasets, MAHNOB-HCI and COHFACE, which reveal outstanding heart rate estimation results.

Palm oil production has been identified as one of the major drivers of deforestation for tropical countries. To meet supply chain objectives, commodity producers and other stakeholders need timely information of land cover dynamics in their supply shed. However, such data are difficult to obtain from suppliers who may lack digital geographic representations of their supply sheds and production locations. Here we present a "community model," a machine learning model trained on pooled data sourced from many different stakeholders, to produce a map of palm probability at global scale. An advantage of this method is the inclusion of varied inputs, the ability to easily update the model as new training data becomes available and run the model on any year that input imagery is available. Inclusion of diverse data sources into one probability map can help establish a shared understanding across stakeholders on the presence and absence of a land cover or commodity (in this case oil palm). The model predictors are annual composites built from publicly available satellite imagery provided by Sentinel-1, Sentinel-2, and ALOS-2, and terrain data from Jaxa (AW3D30) and Copernicus (GLO-30). We provide map outputs as the probability of palm in a given pixel, to reflect the uncertainty of the underlying state (palm or not palm). This version of this model provides global accuracy estimated to be 92% (at 0.5 probability threshold) on an independent test set. This model, and resulting oil palm probability map products are useful for accurately identifying the geographic footprint of palm cultivation. Used in conjunction with timely deforestation information, this palm model is useful for understanding the risk of continued oil palm plantation expansion in sensitive forest areas., Comment: v03

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

This brief shares learnings from the community of Learn and Earn to Achieve Potential (LEAP) organizations and partners as they continue to deepen their practices to enhance the positive impact and durability of programming while also scaling to reach greater numbers of young people in similar situations. After a brief description of the national and local partnering organizations, the first section of the brief explores enhancements that the partners have made to the national models and the consequent development of a set of six core elements of LEAP opportunity pathways that help define what is being scaled.