NeurIPS 2024 Papers

A frequent problem in vision-based reasoning tasks such as object detection and optical character recognition (OCR) is the persistence of specular highlights. Specular highlights appear as bright spots of glare that occur due to the concentrated reflection of light; these spots manifest as image artifacts which occlude computer vision models and are challenging to reconstruct. Despite this, specular highlight removal receives relatively little attention due to the difficulty of acquiring high-quality, real-world data. We introduce a method to generate specular highlight data with near-perfect alignment and present SHDocs—a dataset of specular highlights on document images created using our method. Through our benchmark, we demonstrate that our dataset enables us to surpass the performance of state-of-the-art specular highlight removal models and downstream OCR tasks. We release our dataset, code, and methods publicly to motivate further exploration of image enhancement for practical computer vision challenges.

Compositional Reasoning (CR) entails grasping the significance of attributes, relations, and word order. Recent Vision-Language Models (VLMs), comprising a visual encoder and a Large Language Model (LLM) decoder, have demonstrated remarkable proficiency in such reasoning tasks. This prompts a crucial question: have VLMs effectively tackled the CR challenge? We conjecture that existing CR benchmarks may not adequately push the boundaries of modern VLMs due to the reliance on an LLM only negative text generation pipeline. Consequently, the negatives produced either appear as outliers from the natural language distribution learned by VLMs' LLM decoders or as improbable within the corresponding image context. To address these limitations, we introduce ConMe\footnote{ConMe is an abbreviation for Confuse Me.} -- a compositional reasoning benchmark and a novel data generation pipeline leveraging VLMs to produce `hard CR Q&A'. Through a new concept of VLMs conversing with each other to collaboratively expose their weaknesses, our pipeline autonomously generates, evaluates, and selects challenging compositional reasoning questions, establishing a robust CR benchmark, also subsequently validated manually. Our benchmark provokes a noteworthy, up to 33%, decrease in CR performance compared to preceding benchmarks, reinstating the CR challenge even for state-of-the-art VLMs.

Speech datasets available in the public domain are often underutilized because of challenges in accessibility and interoperability. To address this, a system to survey, catalog, and curate existing speech datasets was developed, enabling reproducible evaluation of automatic speech recognition (ASR) systems. The system was applied to curate over 24 datasets and evaluate 25 ASR models, with a specific focus on Polish. This research represents the most extensive comparison to date of commercial and free ASR systems for the Polish language, drawing insights from 600 system-model-test set evaluations across 8 analysis scenarios. Curated datasets and benchmark results are available publicly. The evaluation tools are open-sourced to support reproducibility of the benchmark, encourage community-driven improvements, and facilitate adaptation for other languages.

Training high-quality deep models necessitates vast amounts of data, resulting in overwhelming computational and memory demands. Recently, data pruning, distillation, and coreset selection have been developed to streamline data volume by \textit{retaining}, \textit{synthesizing}, or \textit{selecting} a small yet informative subset from the full set. Among these methods, data pruning incurs the least additional training cost and offers the most practical acceleration benefits. However, it is the most vulnerable, often suffering significant performance degradation with imbalanced or biased data schema, thus raising concerns about its accuracy and reliability in on-device deployment. Therefore, there is a looming need for a new data pruning paradigm that maintains the efficiency of previous practices while ensuring balance and robustness. Unlike the fields of computer vision and natural language processing, where mature solutions have been developed to address these issues, graph neural networks (GNNs) continue to struggle with increasingly large-scale, imbalanced, and noisy datasets, lacking a unified dataset pruning solution. To achieve this, we introduce a novel dynamic soft-pruning method, \ourmethod, designed to update the training ``basket'' during the process using trainable prototypes. \ourmethod first constructs a well-modeled graph embedding hypersphere and then samples \textit{representative, balanced, and unbiased subsets} from this embedding space, which achieves the goal we called {\fontfamily{lmtt}\selectfont \textbf{Graph Training Debugging}}. Extensive experiments on four datasets across three GNN backbones, demonstrate that \ourmethod (I) achieves or surpasses the performance of the full dataset with $30\%\sim50\%$ fewer training samples, (II) attains up to a $2.81\times$ lossless training speedup, and (III) outperforms state-of-the-art pruning methods in imbalanced training and noisy training scenarios by $0.3\%\sim4.3\%$ and $3.6\%\sim7.8\%$, respectively.

We introduce graphcodes, a novel multi-scale summary of the topological properties of a dataset that is based on the well-established theory of persistent homology. Graphcodes handle datasets that are filtered along two real-valued scale parameters. Such multi-parameter topological summaries are usually based on complicated theoretical foundations and difficult to compute; in contrast, graphcodes yield an informative and interpretable summary and can be computed as efficient as one-parameter summaries. Moreover, a graphcode is simply an embedded graph and can therefore be readily integrated in machine learning pipelines using graph neural networks. We describe such a pipeline and demonstrate that graphcodes achieve better classification accuracy than state-of-the-art approaches on various datasets.

Bayes' rule naturally allows for inference refinement in a streaming fashion, without the need to recompute posteriors from scratch whenever new data arrives. In principle, Bayesian streaming is straightforward: we update our prior with the available data and use the resulting posterior as a prior when processing the next data chunk. In practice, however, this recipe entails i) approximating an intractable posterior at each time step; and ii) encapsulating results appropriately to allow for posterior propagation. For continuous state spaces, variational inference (VI) is particularly convenient due to its scalability and the tractability of variational posteriors, For discrete state spaces, however, state-of-the-art VI results in analytically intractable approximations that are ill-suited for streaming settings. To enable streaming Bayesian inference over discrete parameter spaces, we propose streaming Bayes GFlowNets (abbreviated as SB-GFlowNets) by leveraging the recently proposed GFlowNets --- a powerful class of amortized samplers for discrete compositional objects. Notably, SB-GFlowNet approximates the initial posterior using a standard GFlowNet and subsequently updates it using a tailored procedure that requires only the newly observed data. Our case studies in linear preference learning and phylogenetic inference showcase the effectiveness of SB-GFlowNets in sampling from an unnormalized posterior in a streaming setting. As expected, we also observe that SB-GFlowNets is significantly faster than repeatedly training a GFlowNet from scratch to sample from the full posterior.

Large language models (LLMs) exhibit hallucinations in long-form question-answering tasks across various domains and wide applications. Current hallucination detection and mitigation datasets are limited in domain and size, which struggle to scale due to prohibitive labor costs and insufficient reliability of existing hallucination annotators. To facilitate the scalable oversight of LLM hallucinations, this paper introduces an iterative self-training framework that simultaneously and progressively scales up the annotation dataset and improves the accuracy of the annotator. Based on the Expectation Maximization algorithm, in each iteration, the framework first applies an automatic hallucination annotation pipeline for a scaled dataset and then trains a more accurate annotator on the dataset. This new annotator is adopted in the annotation pipeline for the next iteration. Extensive experimental results demonstrate that the finally obtained hallucination annotator with only 7B parameters surpasses GPT-4 and obtains new state-of-the-art hallucination detection results on HaluEval and HalluQA by zero-shot inference. Such an annotator can not only evaluate the hallucination levels of various LLMs on the large-scale dataset but also help to mitigate the hallucination of LLMs generations, with the Natural Language Inference metric increasing from 25% to 37% on HaluEval.

Computational optimal transport (OT) has received massive interests in the machine learning community, and great advances have been gained in the direction of entropic-regularized OT. The Sinkhorn algorithm, as well as its many improved versions, has become the *de facto* solution to large-scale OT problems. However, most of the existing methods behave like first-order methods, which typically require a large number of iterations to converge. More recently, Newton-type methods using sparsified Hessian matrices have demonstrated promising results on OT computation, but there still remain a lot of unresolved open questions. In this article, we make major new progresses towards this direction: first, we propose a novel Hessian sparsification scheme that promises a strict control of the approximation error; second, based on this sparsification scheme, we develop a *safe* Newton-type method that is guaranteed to avoid singularity in computing the search directions; third, the developed algorithm has a clear implementation for practical use, avoiding most hyperparameter tuning; and remarkably, we provide rigorous global and local convergence analysis of the proposed algorithm, which is lacking in the prior literature. Various numerical experiments are conducted to demonstrate the effectiveness of the proposed algorithm in solving large-scale OT problems.

Antibodies are crucial proteins produced by the immune system to eliminate harmful foreign substances and have become pivotal therapeutic agents for treating human diseases. To accelerate the discovery of antibody therapeutics, there is growing interest in constructing language models using antibody sequences. However, the applicability of pre-trained language models for antibody discovery has not been thoroughly evaluated due to the scarcity of labeled datasets. To overcome these limitations, we introduce AVIDa-SARS-CoV-2, a dataset featuring the antigen-variable domain of heavy chain of heavy chain antibody (VHH) interactions obtained from two alpacas immunized with severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) spike proteins. AVIDa-SARS-CoV-2 includes binary labels indicating the binding or non-binding of diverse VHH sequences to 12 SARS-CoV-2 mutants, such as the Delta and Omicron variants. Furthermore, we release VHHCorpus-2M, a pre-training dataset for antibody language models, containing over two million VHH sequences. We report benchmark results for predicting SARS-CoV-2-VHH binding using VHHBERT pre-trained on VHHCorpus-2M and existing general protein and antibody-specific pre-trained language models. These results confirm that AVIDa-SARS-CoV-2 provides valuable benchmarks for evaluating the representation capabilities of antibody language models for binding prediction, thereby facilitating the development of AI-driven antibody discovery. The datasets are available at https://datasets.cognanous.com.

Despite their remarkable abilities in various tasks, large language models (LLMs) still struggle with real-time information (e.g., new facts and terms) due to the knowledge cutoff in their development process. However, existing benchmarks focus on outdated content and limited fields, facing difficulties in real-time updating and leaving new terms unexplored. To address this problem, we propose an adaptive benchmark, NewTerm, for real-time evaluation of new terms. We design a highly automated construction method to ensure high-quality benchmark construction with minimal human effort, allowing flexible updates for real-time information. Empirical results on various LLMs demonstrate over 20% performance reduction caused by new terms. Additionally, while updates to the knowledge cutoff of LLMs can cover some of the new terms, they are unable to generalize to more distant new terms. We also analyze which types of terms are more challenging and why LLMs struggle with new terms, paving the way for future research. Finally, we construct NewTerm 2022 and 2023 to evaluate the new terms updated each year and will continue updating annually. The benchmark and codes can be found at https://anonymous.4open.science/r/NewTerms.

Traditionally, 3d indoor datasets have generally prioritized scale over ground-truth accuracy in order to obtain improved generalization. However, using these datasets to evaluate dense geometry tasks, such as depth rendering, can be problematic as the meshes of the dataset are often incomplete and may produce wrong ground truth to evaluate the details. In this paper, we propose SCRREAM, a dataset annotation framework that allows annotation of fully dense meshes of objects in the scene and registers camera poses on the real image sequence, which can produce accurate ground truth for both sparse 3D as well as dense 3D tasks. We show the details of the dataset annotation pipeline and showcase four possible variants of datasets that can be obtained from our framework with example scenes, such as indoor reconstruction and SLAM, scene editing \& object removal, human reconstruction and 6d pose estimation. Recent pipelines for indoor reconstruction and SLAM serve as new benchmarks. In contrast to previous indoor dataset, our design allows to evaluate dense geometry tasks on eleven sample scenes against accurately rendered ground truth depth maps.

A recent study has shown that large-scale visual datasets are very biased: they can be easily classified by modern neural networks. However, the concrete forms of bias among these datasets remain unclear. In this study, we propose a framework to identify the unique visual attributes distinguishing these datasets. Our approach applies various transformations to extract semantic, structural, boundary, color, and frequency information from datasets, and assess how much each type of information reflects their bias. We further decompose their semantic bias with object-level analysis, and leverage natural language methods to generate detailed, open-ended descriptions of each dataset's characteristics. Our work aims to help researchers understand the bias in existing large-scale pre-training datasets, and build more diverse and representative ones in the future. Our project page and code are available at boyazeng.github.io/understand_bias.

Understanding how humans interact with each other is key to building realistic multi-human virtual reality systems. This area remains relatively unexplored due to the lack of large-scale datasets. Recent datasets focusing on this issue mainly consist of activities captured entirely in controlled indoor environments with choreographed actions, significantly affecting their diversity. To address this, we introduce Harmony4D, a multi-view video dataset for human-human interaction featuring in-the-wild activities such as wrestling, dancing, MMA, and more. We use a flexible multi-view capture system to record these dynamic activities and provide annotations for human detection, tracking, 2D/3D pose estimation, and mesh recovery for closely interacting subjects. We propose a novel markerless algorithm to track 3D human poses in severe occlusion and close interaction to obtain our annotations with minimal manual intervention. Harmony4D consists of 1.66 million images and 3.32 million human instances from more than 20 synchronized cameras with 208 video sequences spanning diverse environments and 24 unique subjects. We rigorously evaluate existing state-of-the-art methods for mesh recovery and highlight their significant limitations in modeling close interaction scenarios. Additionally, we fine-tune a pre-trained HMR2.0 model on Harmony4D and demonstrate an improved performance of 54.8% PVE in scenes with severe occlusion and contact. “Harmony—a cohesive alignment of human behaviors." Code and data are available at https://jyuntins.github.io/harmony4d/.

Data curation is a field with origins in librarianship and archives, whose scholarship and thinking on data issues go back centuries, if not millennia. The field of machine learning is increasingly observing the importance of data curation to the advancement of both applications and fundamental understanding of machine learning models -- evidenced not least by the creation of the Datasets and Benchmarks track itself. This work provides an analysis of recent dataset development practices at NeurIPS through the lens of data curation. We present an evaluation framework for dataset documentation, consisting of a rubric and toolkit developed through a thorough literature review of data curation principles. We use the framework to systematically assess the strengths and weaknesses in current dataset development practices of 60 datasets published in the NeurIPS Datasets and Benchmarks track from 2021-2023. We summarize key findings and trends. Results indicate greater need for documentation about environmental footprint, ethical considerations, and data management. We suggest targeted strategies and resources to improve documentation in these areas and provide recommendations for the NeurIPS peer-review process that prioritize rigorous data curation in ML. We also provide guidelines for dataset developers on the use of our rubric as a standalone tool. Finally, we provide results in the format of a dataset that showcases aspects of recommended data curation practices. Our rubric and results are of interest for improving data curation practices broadly in the field of ML as well as to data curation and science and technology studies scholars studying practices in ML. Our aim is to support continued improvement in interdisciplinary research on dataset practices, ultimately improving the reusability and reproducibility of new datasets and benchmarks, enabling standardized and informed human oversight, and strengthening the foundation of rigorous and responsible ML research.

Deep neural networks, costly to train and rich in intellectual property value, are increasingly threatened by model extraction attacks that compromise their confiden- tiality. Previous attacks have succeeded in reverse-engineering model parameters up to a precision of float64 for models trained on random data with at most three hidden layers using cryptanalytical techniques. However, the process was identified to be very time consuming and not feasible for larger and deeper models trained on standard benchmarks. Our study evaluates the feasibility of parameter extraction methods of Carlini et al. [1] further enhanced by Canales-Martínez et al. [2] for models trained on standard benchmarks. We introduce a unified codebase that integrates previous methods and reveal that computational tools can significantly influence performance. We develop further optimisations to the end-to-end attack and improve the efficiency of extracting weight signs by up to 14.8 times com- pared to former methods through the identification of easier and harder to extract neurons. Contrary to prior assumptions, we identify extraction of weights, not extraction of weight signs, as the critical bottleneck. With our improvements, a 16,721 parameter model with 2 hidden layers trained on MNIST is extracted within only 98 minutes compared to at least 150 minutes previously. Finally, addressing methodological deficiencies observed in previous studies, we propose new ways of robust benchmarking for future model extraction attacks.

Reflective photoplethysmography (PPG) has become the default sensing technique in wearable devices to monitor cardiac activity via a person’s heart rate (HR). However, PPG-based HR estimates can be substantially impacted by factors such as the wearer’s activities, sensor placement and resulting motion artifacts, as well as environmental characteristics such as temperature and ambient light. These and other factors can significantly impact and decrease HR prediction reliability. In this paper, we show that state-of-the-art HR estimation methods struggle when processing representative data from everyday activities in outdoor environments, likely because they rely on existing datasets that captured controlled conditions. We introduce a novel multimodal dataset and benchmark results for continuous PPG recordings during outdoor activities from 16 participants over 13.5 hours, captured from four wearable sensors, each worn at a different location on the body, totaling 216 hours. Our recordings include accelerometer, temperature, and altitude data, as well as a synchronized Lead I-based electrocardiogram for ground-truth HR references. Participants completed a round trip from Zurich to Jungfraujoch, a tall mountain in Switzerland over the course of one day. The trip included outdoor and indoor activities such as walking, hiking, stair climbing, eating, drinking, and resting at various temperatures and altitudes (up to 3,571 m above sea level) as well as using cars, trains, cable cars, and lifts for transport—all of which impacted participants’ physiological dynamics. We also present a novel method that estimates HR values more robustly in such real-world scenarios than existing baselines. Dataset & code for HR estimation: https://siplab.org/projects/WildPPG

The security threat of backdoor attacks is a central concern for deep neural networks (DNNs). Recently, without poisoned data, unlearning models with clean data and then learning a pruning mask have contributed to backdoor defense. Additionally, vanilla fine-tuning with those clean data can help recover the lost clean accuracy. However, the behavior of clean unlearning is still under-explored, and vanilla fine-tuning unintentionally induces back the backdoor effect. In this work, we first investigate model unlearning from the perspective of weight changes and gradient norms, and find two interesting observations in the backdoored model: 1) the weight changes between poison and clean unlearning are positively correlated, making it possible for us to identify the backdoored-related neurons without using poisoned data; 2) the neurons of the backdoored model are more active (*i.e.*, larger gradient norm) than those in the clean model, suggesting the need to suppress the gradient norm during fine-tuning. Then, we propose an effective two-stage defense method. In the first stage, an efficient *Neuron Weight Change (NWC)-based Backdoor Reinitialization* is proposed based on observation 1). In the second stage, based on observation 2), we design an *Activeness-Aware Fine-Tuning* to replace the vanilla fine-tuning. Extensive experiments, involving eight backdoor attacks on three benchmark datasets, demonstrate the superior performance of our proposed method compared to recent state-of-the-art backdoor defense approaches. The code is available at https://github.com/linweiii/TSBD.git.

Set theory is foundational to mathematics and, when sets are finite, to reasoning about the world. An intelligent system should perform set operations consistently, regardless of superficial variations in the operands. Initially designed for semantically-oriented NLP tasks, large language models (LLMs) are now being evaluated on algorithmic tasks. Because sets are comprised of arbitrary symbols (e.g. numbers, words), they provide an opportunity to test, systematically, the invariance of LLMs’ algorithmic abilities under simple lexical or semantic variations. To this end, we present the SETLEXSEM CHALLENGE, a synthetic benchmark that evaluates the performance of LLMs on set operations. SETLEXSEM assesses the robustness of LLMs’ instruction-following abilities under various conditions, focusing on the set operations and the nature and construction of the set members. Evaluating seven LLMs with SETLEXSEM, we find that they exhibit poor robust- ness to variation in both operation and operands. We show – via the framework’s systematic sampling of set members along lexical and semantic dimensions – that LLMs are not only not robust to variation along these dimensions but demonstrate unique failure modes in particular, easy-to-create semantic groupings of "deceptive" sets. We find that rigorously measuring language model robustness to variation in frequency and length is challenging and present an analysis that measures them in- dependently. The code for reproducing the results of this paper, and for generating the SETLEXSEM CHALLENGE dataset, is available https://github.com/amazon-science/SetLexSem-Challenge.

Recent work has shown the immense potential of synthetically generated datasets for training large language models (LLMs), especially for acquiring targeted skills. Current large-scale math instruction tuning datasets such as MetaMathQA (Yu et al., 2024) and MAmmoTH (Yue et al., 2024) are constructed using outputs from closed-source LLMs with commercially restrictive licenses. A key reason limiting the use of open-source LLMs in these data generation pipelines has been the wide gap between the mathematical skills of the best closed-source LLMs, such as GPT-4, and the best open-source LLMs. Building on the recent progress in open-source LLMs, our proposed prompting novelty, and some brute-force scaling, we construct OpenMathInstruct-1, a math instruction tuning dataset with 1.8M problem-solution pairs. The dataset is constructed by synthesizing code-interpreter solutions for GSM8K and MATH, two popular math reasoning benchmarks, using the recently released and permissively licensed Mixtral model. Our best model, OpenMath-CodeLlama-70B, trained on a subset of OpenMathInstruct-1, achieves a score of 84.6% on GSM8K and 50.7% on MATH, which is competitive with the best gpt-distilled models. We will release our code, models, and the OpenMathInstruct-1 dataset under a commercially permissive license.