Search

Self-Supervised Learning (SSL) has proven to be effective in various domains, including speech processing. However, SSL is computationally and memory expensive. This is in part due the quadratic complexity of multi-head self-attention (MHSA). Alternatives for MHSA have been proposed and used in the speech domain, but have yet to be investigated properly in an SSL setting. In this work, we study the effects of replacing MHSA with recent state-of-the-art alternatives that have linear complexity, namely, HyperMixing, Fastformer, SummaryMixing, and Mamba. We evaluate these methods by looking at the speed, the amount of VRAM consumed, and the performance on the SSL MP3S benchmark. Results show that these linear alternatives maintain competitive performance compared to MHSA while, on average, decreasing VRAM consumption by around 20% to 60% and increasing speed from 7% to 65% for input sequences ranging from 20 to 80 seconds., Comment: Accepted in the IEEE Soken Language Technology Workshop 2024

Large Language Models (LLMs) have shown their ability to improve the performance of speech recognizers by effectively rescoring the n-best hypotheses generated during the beam search process. However, the best way to exploit recent generative instruction-tuned LLMs for hypothesis rescoring is still unclear. This paper proposes a novel method that uses instruction-tuned LLMs to dynamically expand the n-best speech recognition hypotheses with new hypotheses generated through appropriately-prompted LLMs. Specifically, we introduce a new zero-shot method for ASR n-best rescoring, which combines confidence scores, LLM sequence scoring, and prompt-based hypothesis generation. We compare Llama-3-Instruct, GPT-3.5 Turbo, and GPT-4 Turbo as prompt-based generators with Llama-3 as sequence scorer LLM. We evaluated our approach using different speech recognizers and observed significant relative improvement in the word error rate (WER) ranging from 5% to 25%., Comment: IEEE Spoken Language Technology Workshop

Benchmarks are essential in the design of modern HPC installations, as they define key aspects of system components. Beyond synthetic workloads, it is crucial to include real applications that represent user requirements into benchmark suites, to guarantee high usability and widespread adoption of a new system. Given the significant investments in leadership-class supercomputers of the exascale era, this is even more important and necessitates alignment with a vision of Open Science and reproducibility. In this work, we present the JUPITER Benchmark Suite, which incorporates 16 applications from various domains. It was designed for and used in the procurement of JUPITER, the first European exascale supercomputer. We identify requirements and challenges and outline the project and software infrastructure setup. We provide descriptions and scalability studies of selected applications and a set of key takeaways. The JUPITER Benchmark Suite is released as open source software with this work at https://github.com/FZJ-JSC/jubench., Comment: To be published in Proceedings of The International Conference for High Performance Computing Networking, Storage, and Analysis (SC '24) (2024)

Fine-tuning large language models (LLMs) on human preferences, typically through reinforcement learning from human feedback (RLHF), has proven successful in enhancing their capabilities. However, ensuring the safety of LLMs during the fine-tuning remains a critical concern, and mitigating the potential conflicts in safety and helpfulness is costly in RLHF. To address this issue, we propose a supervised learning framework called Bi-Factorial Preference Optimization (BFPO), which re-parameterizes a joint RLHF objective of both safety and helpfulness into a single supervised learning objective. In the supervised optimization, a labeling function is used to capture global preferences ranking to balance both safety and helpfulness. To evaluate BFPO, we develop a benchmark including comprehensive discriminative and generative tasks for helpfulness and harmlessness. The results indicate that our method significantly outperforms existing approaches in both safety and helpfulness. Moreover, BFPO eliminates the need for human prompting and annotation in LLM fine-tuning while achieving the same level of safety as methods that heavily rely on human labor, with less than 10% of the computational resources. The training recipes and models will be released.

Information retrieval across different languages is an increasingly important challenge in natural language processing. Recent approaches based on multilingual pre-trained language models have achieved remarkable success, yet they often optimize for either monolingual, cross-lingual, or multilingual retrieval performance at the expense of others. This paper proposes a novel hybrid batch training strategy to simultaneously improve zero-shot retrieval performance across monolingual, cross-lingual, and multilingual settings while mitigating language bias. The approach fine-tunes multilingual language models using a mix of monolingual and cross-lingual question-answer pair batches sampled based on dataset size. Experiments on XQuAD-R, MLQA-R, and MIRACL benchmark datasets show that the proposed method consistently achieves comparable or superior results in zero-shot retrieval across various languages and retrieval tasks compared to monolingual-only or cross-lingual-only training. Hybrid batch training also substantially reduces language bias in multilingual retrieval compared to monolingual training. These results demonstrate the effectiveness of the proposed approach for learning language-agnostic representations that enable strong zero-shot retrieval performance across diverse languages., Comment: 15 pages, 2 figures, 13 tables

This work provides a comprehensive exploration of various methods in solving incompressible flows using a projection method, and their relation to the occurrence and management of checkerboard oscillations. It employs an algebraic symmetry-preserving framework, clarifying the derivation and implementation of discrete operators while also addressing the associated numerical errors. The lack of a proper definition for the checkerboard problem is addressed by proposing a physics-based coefficient. This coefficient, rooted in the disparity between the compact- and wide-stencil Laplacian operators, is able to quantify oscillatory solution fields with a physics-based, global, normalised, non-dimensional value. The influence of mesh and time-step refinement on the occurrence of checkerboarding is highlighted. Therefore, single measurements using this coefficient should be considered with caution, as the value presents little use without any context and can either suggest mesh refinement or use of a different solver. In addition, an example is given on how to employ this coefficient, by establishing a negative feedback between the level of checkerboarding and the inclusion of a pressure predictor, to dynamically balance the checkerboarding and numerical dissipation. This method is tested for laminar and turbulent flows, demonstrating its capabilities in obtaining this dynamical balance, without requiring user input. The method is able to achieve low numerical dissipation in absence of oscillations or diminish oscillation on skew meshes, while it shows minimal loss in accuracy for a turbulent test case. Despite its advantages, the method exhibits a slight decrease in the second-order relation between time-step size and pressure error, suggesting that other feedback mechanisms could be of interest.

This article is devoted to the analysis of inverse source problems for Stokes systems in unbounded domains where the corresponding velocity flow is observed on a surface. Our main objective is to study the unique determination of general class of time-dependent and vector-valued source terms with potentially unknown divergence. Taking into account the challenges inherent in this class of inverse source problems, we aim to identify the most precise conditions that ensure their resolution. Motivated by various fluid motion problems, we explore several class of boundary measurements. Our proofs are based on different arguments, including unique continuation properties for Stokes systems, approximate controllability, complex analysis, and the application of explicit harmonic functions and explicit solutions for Stokes systems. This analysis is complemented by a reconstruction algorithm and examples of numerical computations.

Federated Learning (FL) in the Internet of Things (IoT) environments can enhance machine learning by utilising decentralised data, but at the same time, it might introduce significant privacy and security concerns due to the constrained nature of IoT devices. This represents a research challenge that we aim to address in this paper. We systematically analysed recent literature to identify privacy threats in FL within IoT environments, and evaluate the defensive measures that can be employed to mitigate these threats. Using a Systematic Literature Review (SLR) approach, we searched five publication databases (Scopus, IEEE Xplore, Wiley, ACM, and Science Direct), collating relevant papers published between 2017 and April 2024, a period which spans from the introduction of FL until now. Guided by the PRISMA protocol, we selected 49 papers to focus our systematic review on. We analysed these papers, paying special attention to the privacy threats and defensive measures -- specifically within the context of IoT -- using inclusion and exclusion criteria tailored to highlight recent advances and critical insights. We identified various privacy threats, including inference attacks, poisoning attacks, and eavesdropping, along with defensive measures such as Differential Privacy and Secure Multi-Party Computation. These defences were evaluated for their effectiveness in protecting privacy without compromising the functional integrity of FL in IoT settings. Our review underscores the necessity for robust and efficient privacy-preserving strategies tailored for IoT environments. Notably, there is a need for strategies against replay, evasion, and model stealing attacks. Exploring lightweight defensive measures and emerging technologies such as blockchain may help improve the privacy of FL in IoT, leading to the creation of FL models that can operate under variable network conditions.

One of us (L.S.) and H. Verlinde independently conjectured a holographic duality between the double-scaled SYK model at infinite temperature and dimensionally reduced $(2+1)$-dimensional de Sitter space [1]-[8]. Beyond the statement that such a duality exists there was deep disagreement between the two proposals [9]. In this note, we trace the origin of the disagreement to a superficial similarity between two q-deformed algebraic structures: the algebra of "chords" in DSSYK, and the algebra of line operators in the Chern-Simons formulation of 3D de Sitter gravity. Assuming that these two structures are the same requires an identification of parameters [7][10] which leads to a collapse of the separation of scales [9] -- the separation being required by the semiclassical limit [3][9]. Dropping that assumption restores the separation of scales but leaves unexplained the relation between chords and Chern-Simons line operators. In this note we point out the existence of a third q-deformed algebra that appears in DSSYK: the algebra of ``wee-chords." Identifying the Chern-Simons line operators with wee-chords removes the discrepancy and leads to a satisfying relation between the two sides of the duality., Comment: 30 + 15 pages, 3 figures

Liver zonation is a fundamental characteristic of hepatocyte spatial heterogeneity, which is challenging to recapitulate in traditional cell cultures. This study presents a novel microfluidic device designed to induce zonation in liver cell cultures by establishing an oxygen gradient using standard laboratory gases. The device consists of two layers; a bottom layer containing a gas channel network that delivers high and low oxygenated gases to create three distinct zones within the cell culture chamber in the layer above. Computational simulations and ratiometric oxygen sensing were employed to validate the oxygen gradient, demonstrating that stable oxygen levels were achieved within two hours. Liver zonation was confirmed using immunofluorescence staining, which showed zonated albumin production in HepG2 cells directly correlating with oxygen levels and mimicking in-vivo zonation behavior. This user-friendly device supports studies on liver zonation and related metabolic disease mechanisms in vitro. It can also be utilized for experiments that necessitate precise gas concentration gradients, such as hypoxia-related research areas focused on angiogenesis and cancer development.

Many social and environmental phenomena are associated with macroscopic changes in the built environment, captured by satellite imagery on a global scale and with daily temporal resolution. While widely used for prediction, these images and especially image sequences remain underutilized for causal inference, especially in the context of randomized controlled trials (RCTs), where causal identification is established by design. In this paper, we develop and compare a set of general tools for analyzing Conditional Average Treatment Effects (CATEs) from temporal satellite data that can be applied to any RCT where geographical identifiers are available. Through a simulation study, we analyze different modeling strategies for estimating CATE in sequences of satellite images. We find that image sequence representation models with more parameters generally yield a greater ability to detect heterogeneity. To explore the role of model and data choice in practice, we apply the approaches to two influential RCTs -- Banerjee et al. (2015), a poverty study in Cusco, Peru, and Bolsen et al. (2014), a water conservation experiment in Georgia, USA. We benchmark our image sequence models against image-only, tabular-only, and combined image-tabular data sources, summarizing practical implications for investigators in a multivariate analysis. Land cover classifications over satellite images facilitate interpretation of what image features drive heterogeneity. We also show robustness to data and model choice of satellite-based generalization of the RCT results to larger geographical areas outside the original. Overall, this paper shows how satellite sequence data can be incorporated into the analysis of RCTs, and provides evidence about the implications of data, model, and evaluation metric choice for causal analysis.

Free-form shape optimization techniques are investigated to improve the separation efficiency of structured packings in laboratory-scale distillation columns. A simplified simulation model based on computational fluid dynamics (CFD) for the mass transfer in the distillation column is used and a corresponding shape optimization problem is formulated. The goal of the optimization is to increase the mass transfer in the column by changing the packing's shape, which has been previously used as criterion for increasing the separation efficiency of the column. The computational shape optimization yields promising results, with an increased mass transfer of nearly 20 %. For validation, the resulting optimized shape is additively manufactured using 3D-printing and investigated experimentally. The experimental results are in good agreement with the performance improvement predicted by the computational model, yielding an increase in separation efficiency of around 20 %.

For medical imaging AI models to be clinically impactful, they must generalize. However, this goal is hindered by (i) diverse types of distribution shifts, such as temporal, demographic, and label shifts, and (ii) limited diversity in datasets that are siloed within single medical institutions. While these limitations have spurred interest in federated learning, current evaluation benchmarks fail to evaluate different shifts simultaneously. However, in real healthcare settings, multiple types of shifts co-exist, yet their impact on medical imaging performance remains unstudied. In response, we introduce FedMedICL, a unified framework and benchmark to holistically evaluate federated medical imaging challenges, simultaneously capturing label, demographic, and temporal distribution shifts. We comprehensively evaluate several popular methods on six diverse medical imaging datasets (totaling 550 GPU hours). Furthermore, we use FedMedICL to simulate COVID-19 propagation across hospitals and evaluate whether methods can adapt to pandemic changes in disease prevalence. We find that a simple batch balancing technique surpasses advanced methods in average performance across FedMedICL experiments. This finding questions the applicability of results from previous, narrow benchmarks in real-world medical settings., Comment: Accepted at MICCAI 2024. Code is available at: https://github.com/m1k2zoo/FedMedICL

Crowd counting is gaining societal relevance, particularly in domains of Urban Planning, Crowd Management, and Public Safety. This paper introduces Fourier-guided attention (FGA), a novel attention mechanism for crowd count estimation designed to address the inefficient full-scale global pattern capture in existing works on convolution-based attention networks. FGA efficiently captures multi-scale information, including full-scale global patterns, by utilizing Fast-Fourier Transformations (FFT) along with spatial attention for global features and convolutions with channel-wise attention for semi-global and local features. The architecture of FGA involves a dual-path approach: (1) a path for processing full-scale global features through FFT, allowing for efficient extraction of information in the frequency domain, and (2) a path for processing remaining feature maps for semi-global and local features using traditional convolutions and channel-wise attention. This dual-path architecture enables FGA to seamlessly integrate frequency and spatial information, enhancing its ability to capture diverse crowd patterns. We apply FGA in the last layers of two popular crowd-counting works, CSRNet and CANNet, to evaluate the module's performance on benchmark datasets such as ShanghaiTech-A, ShanghaiTech-B, UCF-CC-50, and JHU++ crowd. The experiments demonstrate a notable improvement across all datasets based on Mean-Squared-Error (MSE) and Mean-Absolute-Error (MAE) metrics, showing comparable performance to recent state-of-the-art methods. Additionally, we illustrate the interpretability using qualitative analysis, leveraging Grad-CAM heatmaps, to show the effectiveness of FGA in capturing crowd patterns., Comment: Accepted to IJCNN'24

Divergence constraints are present in the governing equations of numerous physical phenomena, and they usually lead to a Poisson equation whose solution represents a bottleneck in many simulation codes. Algebraic Multigrid (AMG) is arguably the most powerful preconditioner for Poisson's equation, and its effectiveness results from the complementary roles played by the smoother, responsible for damping high-frequency error components, and the coarse-grid correction, which in turn reduces low-frequency modes. This work presents several strategies to make AMG more compute-intensive by leveraging reflection, translational and rotational symmetries. AMGR, our final proposal, does not require boundary conditions to be symmetric, therefore applying to a broad range of academic and industrial configurations. It is based on a multigrid reduction framework that introduces an aggressive coarsening to the multigrid hierarchy, reducing the memory footprint, setup and application costs of the top-level smoother. While preserving AMG's excellent convergence, AMGR allows replacing the standard sparse matrix-vector product with the more compute-intensive sparse matrix-matrix product, yielding significant accelerations. Numerical experiments on industrial CFD applications demonstrated up to 70% speed-ups when solving Poisson's equation with AMGR instead of AMG. Additionally, strong and weak scalability analyses revealed no significant degradation.

Driven by advancements in sensing and computing, deep reinforcement learning (DRL)-based methods have demonstrated significant potential in effectively tackling distribution system restoration (DSR) challenges under uncertain operational scenarios. However, the data-intensive nature of DRL poses obstacles in achieving satisfactory DSR solutions for large-scale, complex distribution systems. Inspired by the transformative impact of emerging foundation models, including large language models (LLMs), across various domains, this paper explores an innovative approach harnessing LLMs' powerful computing capabilities to address scalability challenges inherent in conventional DRL methods for solving DSR. To our knowledge, this study represents the first exploration of foundation models, including LLMs, in revolutionizing conventional DRL applications in power system operations. Our contributions are twofold: 1) introducing a novel LLM-powered Physics-Informed Decision Transformer (PIDT) framework that leverages LLMs to transform conventional DRL methods for DSR operations, and 2) conducting comparative studies to assess the performance of the proposed LLM-powered PIDT framework at its initial development stage for solving DSR problems. While our primary focus in this paper is on DSR operations, the proposed PIDT framework can be generalized to optimize sequential decision-making across various power system operations.

SpeechBrain is an open-source Conversational AI toolkit based on PyTorch, focused particularly on speech processing tasks such as speech recognition, speech enhancement, speaker recognition, text-to-speech, and much more. It promotes transparency and replicability by releasing both the pre-trained models and the complete "recipes" of code and algorithms required for training them. This paper presents SpeechBrain 1.0, a significant milestone in the evolution of the toolkit, which now has over 200 recipes for speech, audio, and language processing tasks, and more than 100 models available on Hugging Face. SpeechBrain 1.0 introduces new technologies to support diverse learning modalities, Large Language Model (LLM) integration, and advanced decoding strategies, along with novel models, tasks, and modalities. It also includes a new benchmark repository, offering researchers a unified platform for evaluating models across diverse tasks., Comment: Submitted to JMLR (Machine Learning Open Source Software)

In real-world environments, continual learning is essential for machine learning models, as they need to acquire new knowledge incrementally without forgetting what they have already learned. While pretrained language models have shown impressive capabilities on various static tasks, applying them to continual learning poses significant challenges, including avoiding catastrophic forgetting, facilitating knowledge transfer, and maintaining parameter efficiency. In this paper, we introduce MoCL-P, a novel lightweight continual learning method that addresses these challenges simultaneously. Unlike traditional approaches that continuously expand parameters for newly arriving tasks, MoCL-P integrates task representation-guided module composition with adaptive pruning, effectively balancing knowledge integration and computational overhead. Our evaluation across three continual learning benchmarks with up to 176 tasks shows that MoCL-P achieves state-of-the-art performance and improves parameter efficiency by up to three times, demonstrating its potential for practical applications where resource requirements are constrained.

We demonstrate a free-space amplitude modulator for mid-infrared radiation (lambda=9.6 um) that operates at room temperature up to at least 20 GHz (above the -3dB cutoff frequency measured at 8.2 GHz). The device relies on the ultra-fast transition between weak and strong-coupling regimes induced by the variation of the applied bias voltage. Such transition induces a modulation of the device reflectivity. It is made of a semiconductor heterostructure enclosed in a judiciously designed array of metal-metal optical resonators, that - all-together - behave as an electrically tunable surface. At negative bias, it operates in the weak light-matter coupling regime. Upon application of an appropriate positive bias, the quantum wells populate with electrons and the device transitions to the strong-coupling regime. The modulator transmission keeps linear with input RF power in the 0dBm - 9dBm range. The increase of optical powers up to 25 mW exhibit a weak beginning saturation a little bit below.

Merging Large Language Models (LLMs) is a cost-effective technique for combining multiple expert LLMs into a single versatile model, retaining the expertise of the original ones. However, current approaches often overlook the importance of safety alignment during merging, leading to highly misaligned models. This work investigates the effects of model merging on alignment. We evaluate several popular model merging techniques, demonstrating that existing methods do not only transfer domain expertise but also propagate misalignment. We propose a simple two-step approach to address this problem: (i) generating synthetic safety and domain-specific data, and (ii) incorporating these generated data into the optimization process of existing data-aware model merging techniques. This allows us to treat alignment as a skill that can be maximized in the resulting merged LLM. Our experiments illustrate the effectiveness of integrating alignment-related data during merging, resulting in models that excel in both domain expertise and alignment., Comment: Under review

The performance of a model trained with \textit{noisy labels} is often improved by simply \textit{retraining} the model with its own predicted \textit{hard} labels (i.e., $1$/$0$ labels). Yet, a detailed theoretical characterization of this phenomenon is lacking. In this paper, we theoretically analyze retraining in a linearly separable setting with randomly corrupted labels given to us and prove that retraining can improve the population accuracy obtained by initially training with the given (noisy) labels. To the best of our knowledge, this is the first such theoretical result. Retraining finds application in improving training with label differential privacy (DP) which involves training with noisy labels. We empirically show that retraining selectively on the samples for which the predicted label matches the given label significantly improves label DP training at \textit{no extra privacy cost}; we call this \textit{consensus-based retraining}. For e.g., when training ResNet-18 on CIFAR-100 with $\epsilon=3$ label DP, we obtain $6.4\%$ improvement in accuracy with consensus-based retraining.

We review research at the National Physical Laboratory (NPL) in the area of trustworthy artificial intelligence (TAI), and more specifically trustworthy machine learning (TML), in the context of metrology, the science of measurement. We describe three broad themes of TAI: technical, socio-technical and social, which play key roles in ensuring that the developed models are trustworthy and can be relied upon to make responsible decisions. From a metrology perspective we emphasise uncertainty quantification (UQ), and its importance within the framework of TAI to enhance transparency and trust in the outputs of AI systems. We then discuss three research areas within TAI that we are working on at NPL, and examine the certification of AI systems in terms of adherence to the characteristics of TAI.

Fine-tuning large language models on small, high-quality datasets can enhance their performance on specific downstream tasks. Recent research shows that fine-tuning on benign, instruction-following data can inadvertently undo the safety alignment process and increase a model's propensity to comply with harmful queries. Although critical, understanding and mitigating safety risks in well-defined tasks remains distinct from the instruction-following context due to structural differences in the data. Our work addresses the gap in our understanding of these risks across diverse types of data in closed models - where providers control how user data is utilized in the fine-tuning process. We demonstrate how malicious actors can subtly manipulate the structure of almost any task-specific dataset to foster significantly more dangerous model behaviors, while maintaining an appearance of innocuity and reasonable downstream task performance. To address this issue, we propose a novel mitigation strategy that mixes in safety data which mimics the task format and prompting style of the user data, showing this is more effective than existing baselines at re-establishing safety alignment while maintaining similar task performance.

Music evokes emotion in many people. We introduce a novel way to manipulate the emotional content of a song using AI tools. Our goal is to achieve the desired emotion while leaving the original melody as intact as possible. For this, we create an interactive pipeline capable of shifting an input song into a diametrically opposed emotion and visualize this result through Russel's Circumplex model. Our approach is a proof-of-concept for Semantic Manipulation of Music, a novel field aimed at modifying the emotional content of existing music. We design a deep learning model able to assess the accuracy of our modifications to key, SoundFont instrumentation, and other musical features. The accuracy of our model is in-line with the current state of the art techniques on the 4Q Emotion dataset. With further refinement, this research may contribute to on-demand custom music generation, the automated remixing of existing work, and music playlists tuned for emotional progression.

We present FaithFill, a diffusion-based inpainting object completion approach for realistic generation of missing object parts. Typically, multiple reference images are needed to achieve such realistic generation, otherwise the generation would not faithfully preserve shape, texture, color, and background. In this work, we propose a pipeline that utilizes only a single input reference image -having varying lighting, background, object pose, and/or viewpoint. The singular reference image is used to generate multiple views of the object to be inpainted. We demonstrate that FaithFill produces faithful generation of the object's missing parts, together with background/scene preservation, from a single reference image. This is demonstrated through standard similarity metrics, human judgement, and GPT evaluation. Our results are presented on the DreamBooth dataset, and a novel proposed dataset.

In the rapidly evolving landscape of spoken question-answering (SQA), the integration of large language models (LLMs) has emerged as a transformative development. Conventional approaches often entail the use of separate models for question audio transcription and answer selection, resulting in significant resource utilization and error accumulation. To tackle these challenges, we explore the effectiveness of end-to-end (E2E) methodologies for SQA in the medical domain. Our study introduces a novel zero-shot SQA approach, compared to traditional cascade systems. Through a comprehensive evaluation conducted on a new open benchmark of 8 medical tasks and 48 hours of synthetic audio, we demonstrate that our approach requires up to 14.7 times fewer resources than a combined 1.3B parameters LLM with a 1.55B parameters ASR model while improving average accuracy by 0.5\%. These findings underscore the potential of E2E methodologies for SQA in resource-constrained contexts., Comment: Accepted to INTERSPEECH 2024

Relieving the reliance of neural network training on a global back-propagation (BP) has emerged as a notable research topic due to the biological implausibility and huge memory consumption caused by BP. Among the existing solutions, local learning optimizes gradient-isolated modules of a neural network with local errors and has been proved to be effective even on large-scale datasets. However, the reconciliation among local errors has never been investigated. In this paper, we first theoretically study non-greedy layer-wise training and show that the convergence cannot be assured when the local gradient in a module w.r.t. its input is not reconciled with the local gradient in the previous module w.r.t. its output. Inspired by the theoretical result, we further propose a local training strategy that successively regularizes the gradient reconciliation between neighboring modules without breaking gradient isolation or introducing any learnable parameters. Our method can be integrated into both local-BP and BP-free settings. In experiments, we achieve significant performance improvements compared to previous methods. Particularly, our method for CNN and Transformer architectures on ImageNet is able to attain a competitive performance with global BP, saving more than 40% memory consumption., Comment: ICML 2024

Zero-shot and in-context learning enable solving tasks without model fine-tuning, making them essential for developing generative model solutions. Therefore, it is crucial to understand whether a pretrained model can be prompted to approximate any function, i.e., whether it is a universal in-context approximator. While it was recently shown that transformer models do possess this property, these results rely on their attention mechanism. Hence, these findings do not apply to fully recurrent architectures like RNNs, LSTMs, and the increasingly popular SSMs. We demonstrate that RNNs, LSTMs, GRUs, Linear RNNs, and linear gated architectures such as Mamba and Hawk/Griffin can also serve as universal in-context approximators. To streamline our argument, we introduce a programming language called LSRL that compiles to these fully recurrent architectures. LSRL may be of independent interest for further studies of fully recurrent models, such as constructing interpretability benchmarks. We also study the role of multiplicative gating and observe that architectures incorporating such gating (e.g., LSTMs, GRUs, Hawk/Griffin) can implement certain operations more stably, making them more viable candidates for practical in-context universal approximation.

Training Large Language Models (LLMs) relies heavily on distributed implementations, employing multiple GPUs to compute stochastic gradients on model replicas in parallel. However, synchronizing gradients in data parallel settings induces a communication overhead increasing with the number of distributed workers, which can impede the efficiency gains of parallelization. To address this challenge, optimization algorithms reducing inter-worker communication have emerged, such as local optimization methods used in Federated Learning. While effective in minimizing communication overhead, these methods incur significant memory costs, hindering scalability: in addition to extra momentum variables, if communications are only allowed between multiple local optimization steps, then the optimizer's states cannot be sharded among workers. In response, we propose $\textbf{AC}$cumulate while $\textbf{CO}$mmunicate ($\texttt{ACCO}$), a memory-efficient optimization algorithm tailored for distributed training of LLMs. $\texttt{ACCO}$ allows to shard optimizer states across workers, overlaps gradient computations and communications to conceal communication costs, and accommodates heterogeneous hardware. Our method relies on a novel technique to mitigate the one-step delay inherent in parallel execution of gradient computations and communications, eliminating the need for warmup steps and aligning with the training dynamics of standard distributed optimization while converging faster in terms of wall-clock time. We demonstrate the effectiveness of $\texttt{ACCO}$ on several LLMs training and fine-tuning tasks.

We consider a weakly supervised learning problem called Learning from Label Proportions (LLP), where examples are grouped into ``bags'' and only the average label within each bag is revealed to the learner. We study various learning rules for LLP that achieve PAC learning guarantees for classification loss. We establish that the classical Empirical Proportional Risk Minimization (EPRM) learning rule (Yu et al., 2014) achieves fast rates under realizability, but EPRM and similar proportion matching learning rules can fail in the agnostic setting. We also show that (1) a debiased proportional square loss, as well as (2) a recently proposed EasyLLP learning rule (Busa-Fekete et al., 2023) both achieve ``optimistic rates'' (Panchenko, 2002); in both the realizable and agnostic settings, their sample complexity is optimal (up to log factors) in terms of $\epsilon, \delta$, and VC dimension., Comment: Accepted to COLT 2024. Comments welcome

Earth observation data such as satellite imagery can, when combined with machine learning, can have far-reaching impacts on our understanding of the geography of poverty through the prediction of living conditions, especially where government-derived economic indicators are either unavailable or potentially untrustworthy. Recent work has progressed in using Earth Observation (EO) data not only to predict spatial economic outcomes but also to explore cause and effect, an understanding which is critical for downstream policy analysis. In this review, we first document the growth of interest in using satellite images together with EO data in causal analysis. We then trace the relationship between spatial statistics and machine learning methods before discussing four ways in which EO data has been used in causal machine learning pipelines -- (1.) poverty outcome imputation for downstream causal analysis, (2.) EO image deconfounding, (3.) EO-based treatment effect heterogeneity, and (4.) EO-based transportability analysis. We conclude by providing a step-by-step workflow for how researchers can incorporate EO data in causal ML analysis going forward, outlining major choices of data, models, and evaluation metrics., Comment: For a full list of the papers found in the quantitative literature search, see https://github.com/AIandGlobalDevelopmentLab/eo-poverty-review

The motion of rockets is part of the study devoted to the motion of variable mass systems. Notably those in which the mass leaves permanently the considered system. Rockets are propelled forward by the reaction force produced by the hot exhausted gases ejected from their tales in the rearward direction. Thus their motion should not violate Newton's third principle of the equality of action and reaction forces during the exhaustion process. Nor should it violate Newton's second law of motion judged by inertial observers. However a close examination of the study of the motion of rockets in a major part of physics textbooks, if not all reveals erroneous determination of the expression of the thrust force that pushes the rocket in the forward direction. The false expression of the thrust force entails a bad effect on obtaining the right differential equation that governs the motion of rockets. This trap induced some prominent physics authors to pretend the in applicability of Newton's second law in such particular cases. Not only that, but they also modified Newton's second law in order to fit their purposes of obtaining the right differential equation, historically known under the name Tsiolkovsky rocket equation. The object of this paper is to give the true expression of the thrust force and to write the differential equation of motion of rockets without any necessity of modification of the classical laws. The paper also delves into the expression of the change of velocity of rockets, and proves that their motion is uniformly accelerated in the early stages of liftoff from the ground at condition of constant velocity of expulsion of hot gases from their nozzles., Comment: no comments

The performance of deep neural networks is enhanced by ensemble methods, which average the output of several models. However, this comes at an increased cost at inference. Weight averaging methods aim at balancing the generalization of ensembling and the inference speed of a single model by averaging the parameters of an ensemble of models. Yet, naive averaging results in poor performance as models converge to different loss basins, and aligning the models to improve the performance of the average is challenging. Alternatively, inspired by distributed training, methods like DART and PAPA have been proposed to train several models in parallel such that they will end up in the same basin, resulting in good averaging accuracy. However, these methods either compromise ensembling accuracy or demand significant communication between models during training. In this paper, we introduce WASH, a novel distributed method for training model ensembles for weight averaging that achieves state-of-the-art image classification accuracy. WASH maintains models within the same basin by randomly shuffling a small percentage of weights during training, resulting in diverse models and lower communication costs compared to standard parameter averaging methods.

Field-captured video allows for detailed studies of spatiotemporal aspects of animal locomotion, decision-making, and environmental interactions. However, despite the affordability of data capture with mass-produced hardware, storage, processing, and transmission overheads pose a significant hurdle to acquiring high-resolution video from field-deployed camera traps. Therefore, efficient compression algorithms are crucial for monitoring with camera traps that have limited access to power, storage, and bandwidth. In this article, we introduce a new motion analysis-based video compression algorithm designed to run on camera trap devices. We implemented and tested this algorithm using a case study of insect-pollinator motion tracking. The algorithm identifies and stores only image regions depicting motion relevant to pollination monitoring, reducing the overall data size by an average of 84% across a diverse set of test datasets while retaining the information necessary for relevant behavioural analysis. The methods outlined in this paper facilitate the broader application of computer vision-enabled, low-powered camera trap devices for remote, in-situ video-based animal motion monitoring., Comment: 6 pages, 3 figures, 3 tables

Since neural classifiers are known to be sensitive to adversarial perturbations that alter their accuracy, \textit{certification methods} have been developed to provide provable guarantees on the insensitivity of their predictions to such perturbations. Furthermore, in safety-critical applications, the frequentist interpretation of the confidence of a classifier (also known as model calibration) can be of utmost importance. This property can be measured via the Brier score or the expected calibration error. We show that attacks can significantly harm calibration, and thus propose certified calibration as worst-case bounds on calibration under adversarial perturbations. Specifically, we produce analytic bounds for the Brier score and approximate bounds via the solution of a mixed-integer program on the expected calibration error. Finally, we propose novel calibration attacks and demonstrate how they can improve model calibration through \textit{adversarial calibration training}., Comment: 11 pages main paper, appendix included

Cloud native solutions are widely applied in various fields, placing higher demands on the efficient management and utilization of resource platforms. To achieve the efficiency, load forecasting and elastic scaling have become crucial technologies for dynamically adjusting cloud resources to meet user demands and minimizing resource waste. However, existing prediction-based methods lack comprehensive analysis and integration of load characteristics across different time scales. For instance, long-term trend analysis helps reveal long-term changes in load and resource demand, thereby supporting proactive resource allocation over longer periods, while short-term volatility analysis can examine short-term fluctuations in load and resource demand, providing support for real-time scheduling and rapid response. In response to this, our research introduces TempoScale, which aims to enhance the comprehensive understanding of temporal variations in cloud workloads, enabling more intelligent and adaptive decision-making for elastic scaling. TempoScale utilizes the Complete Ensemble Empirical Mode Decomposition with Adaptive Noise algorithm to decompose time-series load data into multiple Intrinsic Mode Functions (IMF) and a Residual Component (RC). First, we integrate the IMF, which represents both long-term trends and short-term fluctuations, into the time series prediction model to obtain intermediate results. Then, these intermediate results, along with the RC, are transferred into a fully connected layer to obtain the final result. Finally, this result is fed into the resource management system based on Kubernetes for resource scaling. Our proposed approach can reduce the Mean Square Error by 5.80% to 30.43% compared to the baselines, and reduce the average response time by 5.58% to 31.15%., Comment: 11pages, 11 figures, 4 tables

Creating Automatic Speech Recognition (ASR) systems that are robust and resilient to classroom conditions is paramount to the development of AI tools to aid teachers and students. In this work, we study the efficacy of continued pretraining (CPT) in adapting Wav2vec2.0 to the classroom domain. We show that CPT is a powerful tool in that regard and reduces the Word Error Rate (WER) of Wav2vec2.0-based models by upwards of 10%. More specifically, CPT improves the model's robustness to different noises, microphones, classroom conditions as well as classroom demographics. Our CPT models show improved ability to generalize to different demographics unseen in the labeled finetuning data.

Applications of Generative AI (Gen AI) are expected to revolutionize a number of different areas, ranging from science & medicine to education. The potential for these seismic changes has triggered a lively debate about the potential risks of the technology, and resulted in calls for tighter regulation, in particular from some of the major tech companies who are leading in AI development. This regulation is likely to put at risk the budding field of open-source generative AI. Using a three-stage framework for Gen AI development (near, mid and long-term), we analyze the risks and opportunities of open-source generative AI models with similar capabilities to the ones currently available (near to mid-term) and with greater capabilities (long-term). We argue that, overall, the benefits of open-source Gen AI outweigh its risks. As such, we encourage the open sourcing of models, training and evaluation data, and provide a set of recommendations and best practices for managing risks associated with open-source generative AI., Comment: Extension of arXiv:2404.17047

Table Question Answering (TQA) aims at composing an answer to a question based on tabular data. While prior research has shown that TQA models lack robustness, understanding the underlying cause and nature of this issue remains predominantly unclear, posing a significant obstacle to the development of robust TQA systems. In this paper, we formalize three major desiderata for a fine-grained evaluation of robustness of TQA systems. They should (i) answer questions regardless of alterations in table structure, (ii) base their responses on the content of relevant cells rather than on biases, and (iii) demonstrate robust numerical reasoning capabilities. To investigate these aspects, we create and publish a novel TQA evaluation benchmark in English. Our extensive experimental analysis reveals that none of the examined state-of-the-art TQA systems consistently excels in these three aspects. Our benchmark is a crucial instrument for monitoring the behavior of TQA systems and paves the way for the development of robust TQA systems. We release our benchmark publicly., Comment: Accepted at NAACL 2024

In the next few years, applications of Generative AI are expected to revolutionize a number of different areas, ranging from science & medicine to education. The potential for these seismic changes has triggered a lively debate about potential risks and resulted in calls for tighter regulation, in particular from some of the major tech companies who are leading in AI development. This regulation is likely to put at risk the budding field of open-source Generative AI. We argue for the responsible open sourcing of generative AI models in the near and medium term. To set the stage, we first introduce an AI openness taxonomy system and apply it to 40 current large language models. We then outline differential benefits and risks of open versus closed source AI and present potential risk mitigation, ranging from best practices to calls for technical and scientific contributions. We hope that this report will add a much needed missing voice to the current public discourse on near to mid-term AI safety and other societal impact., Comment: Accepted to ICML'24 as a position paper

We investigate the spin torque in the topological phase of the 3Q antiferromagnetic (AFM) configuration. We first obtain the band structure to identify the topology and nature of the different gaps of the system and then calculate the spin density in the whole system. We demonstrate the presence of a non-vanishing spin-torque component at the edge of the system and analyze its effect on the texture. Moreover, we show that the direction of the torque depends hugely on the way we cut the system to define the edge.

We propose and study a realistic Continual Learning (CL) setting where learning algorithms are granted a restricted computational budget per time step while training. We apply this setting to large-scale semi-supervised Continual Learning scenarios with sparse label rates. Previous proficient CL methods perform very poorly in this challenging setting. Overfitting to the sparse labeled data and insufficient computational budget are the two main culprits for such a poor performance. Our new setting encourages learning methods to effectively and efficiently utilize the unlabeled data during training. To that end, we propose a simple but highly effective baseline, DietCL, which utilizes both unlabeled and labeled data jointly. DietCL meticulously allocates computational budget for both types of data. We validate our baseline, at scale, on several datasets, e.g., CLOC, ImageNet10K, and CGLM, under constraint budget setups. DietCL outperforms, by a large margin, all existing supervised CL algorithms as well as more recent continual semi-supervised methods. Our extensive analysis and ablations demonstrate that DietCL is stable under a full spectrum of label sparsity, computational budget, and various other ablations.

Fault diagnosis in marine diesel engines is vital for maritime safety and operational efficiency.These engines are integral to marine vessels, and their reliable performance is crucial for safenavigation. Swift identification and resolution of faults are essential to prevent breakdowns,enhance safety, and reduce the risk of catastrophic failures at sea. Proactive fault diagnosisfacilitates timely maintenance, minimizes downtime, and ensures the overall reliability andlongevity of marine diesel engines. This paper explores the importance of fault diagnosis,emphasizing subsystems, common faults, and recent advancements in data-driven approachesfor effective marine diesel engine maintenance

We define and study a long-range version of the XX model, arising as the free-fermion point of the XXZ-type Haldane--Shastry (HS) chain. It has a description via non-unitary fermions, based on the free-fermion Temperley--Lieb algebra, and may also be viewed as an alternating $\mathfrak{gl}(1|1)$ spin chain. Even and odd length behave very differently; we focus on odd length. The model is integrable, and we explicitly identify two commuting hamiltonians. While non-unitary, their spectrum is real by PT-symmetry. One hamiltonian is chiral and quadratic in fermions, while the other is parity-invariant and quartic. Their one-particle spectra have two linear branches, realising a massless relativistic dispersion on the lattice. The appropriate fermionic modes arise from 'quasi-translation' symmetry, which replaces ordinary translation symmetry. The model exhibits exclusion statistics, like the isotropic HS chain, with even more 'extended symmetry' and larger degeneracies., Comment: v1: 6+23 pages, 3 figs, 1 table. v2: various minor improvements; 6+27 pages, 1+3 figs, 1 table

Web-crawled pretraining datasets underlie the impressive "zero-shot" evaluation performance of multimodal models, such as CLIP for classification/retrieval and Stable-Diffusion for image generation. However, it is unclear how meaningful the notion of "zero-shot" generalization is for such multimodal models, as it is not known to what extent their pretraining datasets encompass the downstream concepts targeted for during "zero-shot" evaluation. In this work, we ask: How is the performance of multimodal models on downstream concepts influenced by the frequency of these concepts in their pretraining datasets? We comprehensively investigate this question across 34 models and five standard pretraining datasets (CC-3M, CC-12M, YFCC-15M, LAION-400M, LAION-Aesthetics), generating over 300GB of data artifacts. We consistently find that, far from exhibiting "zero-shot" generalization, multimodal models require exponentially more data to achieve linear improvements in downstream "zero-shot" performance, following a sample inefficient log-linear scaling trend. This trend persists even when controlling for sample-level similarity between pretraining and downstream datasets, and testing on purely synthetic data distributions. Furthermore, upon benchmarking models on long-tailed data sampled based on our analysis, we demonstrate that multimodal models across the board perform poorly. We contribute this long-tail test set as the "Let it Wag!" benchmark to further research in this direction. Taken together, our study reveals an exponential need for training data which implies that the key to "zero-shot" generalization capabilities under large-scale training paradigms remains to be found., Comment: Extended version of the short paper accepted at DPFM, ICLR'24