NeurIPS 2023 Papers

Tracking an arbitrary moving target in a video sequence is the foundation for high-level tasks like video understanding. Although existing visual-based trackers have demonstrated good tracking capabilities in short video sequences, they always perform poorly in complex environments, as represented by the recently proposed global instance tracking task, which consists of longer videos with more complicated narrative content. Recently, several works have introduced natural language into object tracking, desiring to address the limitations of relying only on a single visual modality. However, these selected videos are still short sequences with uncomplicated spatio-temporal and causal relationships, and the provided semantic descriptions are too simple to characterize video content. To address these issues, we (1) first propose a new multi-modal global instance tracking benchmark named MGIT. It consists of 150 long video sequences with a total of 2.03 million frames, aiming to fully represent the complex spatio-temporal and causal relationships coupled in longer narrative content. (2) Each video sequence is annotated with three semantic grains (i.e., action, activity, and story) to model the progressive process of human cognition. We expect this multi-granular annotation strategy can provide a favorable environment for multi-modal object tracking research and long video understanding. (3) Besides, we execute comparative experiments on existing multi-modal object tracking benchmarks, which not only explore the impact of different annotation methods, but also validate that our annotation method is a feasible solution for coupling human understanding into semantic labels. (4) Additionally, we conduct detailed experimental analyses on MGIT, and hope the explored performance bottlenecks of existing algorithms can support further research in multi-modal object tracking. The proposed benchmark, experimental results, and toolkit will be released gradually on http://videocube.aitestunion.com/.

Recently, commonsense reasoning in text generation has attracted much attention. Generative commonsense reasoning is the task that requires machines, given a group of keywords, to compose a single coherent sentence with commonsense plausibility. While existing datasets targeting generative commonsense reasoning focus on everyday scenarios, it is unclear how well machines reason under specific geographical and temporal contexts. We formalize this challenging task as SituatedGen, where machines with commonsense should generate a pair of contrastive sentences given a group of keywords including geographical or temporal entities. We introduce a corresponding English dataset consisting of 8,268 contrastive sentence pairs, which are built upon several existing commonsense reasoning benchmarks with minimal manual labor. Experiments show that state-of-the-art generative language models struggle to generate sentences with commonsense plausibility and still lag far behind human performance. Our dataset is publicly available at https://github.com/yunx-z/situated_gen.

We prove that, for the fundamental regression task of learning a single neuron, training a one-hidden layer ReLU network of any width by gradient flow from a small initialisation converges to zero loss and is implicitly biased to minimise the rank of network parameters. By assuming that the training points are correlated with the teacher neuron, we complement previous work that considered orthogonal datasets. Our results are based on a detailed non-asymptotic analysis of the dynamics of each hidden neuron throughout the training. We also show and characterise a surprising distinction in this setting between interpolator networks of minimal rank and those of minimal Euclidean norm. Finally we perform a range of numerical experiments, which corroborate our theoretical findings.

Learning curve extrapolation aims to predict model performance in later epochs of training, based on the performance in earlier epochs. In this work, we argue that, while the inherent uncertainty in the extrapolation of learning curves warrants a Bayesian approach, existing methods are (i) overly restrictive, and/or (ii) computationally expensive. We describe the first application of prior-data fitted neural networks (PFNs) in this context. A PFN is a transformer, pre-trained on data generated from a prior, to perform approximate Bayesian inference in a single forward pass. We propose LC-PFN, a PFN trained to extrapolate 10 million artificial right-censored learning curves generated from a parametric prior proposed in prior art using MCMC. We demonstrate that LC-PFN can approximate the posterior predictive distribution more accurately than MCMC, while being over 10 000 times faster. We also show that the same LC-PFN achieves competitive performance extrapolating a total of 20 000 real learning curves from four learning curve benchmarks (LCBench, NAS-Bench-201, Taskset, and PD1) that stem from training a wide range of model architectures (MLPs, CNNs, RNNs, and Transformers) on 53 different datasets with varying input modalities (tabular, image, text, and protein data). Finally, we investigate its potential in the context of model selection and find that a simple LC-PFN based predictive early stopping criterion obtains 2 - 6x speed-ups on 45 of these datasets, at virtually no overhead.

Pruning at initialization (PaI) aims to remove weights of neural networks before training in pursuit of training efficiency besides the inference. While off-the-shelf PaI methods manage to find trainable subnetworks that outperform random pruning, their performance in terms of both accuracy and computational reduction is far from satisfactory compared to post-training pruning and the understanding of PaI is missing. For instance, recent studies show that existing PaI methods only able to find good layerwise sparsities not weights, as the discovered subnetworks are surprisingly resilient against layerwise random mask shuffling and weight re-initialization. In this paper, we study PaI from a brand-new perspective -- the topology of subnetworks. In particular, we propose a principled framework for analyzing the performance of Pruning and Initialization (PaI) methods with two quantities, namely, the number of effective paths and effective nodes. These quantities allow for a more comprehensive understanding of PaI methods, giving us an accurate assessment of different subnetworks at initialization. We systematically analyze the behavior of various PaI methods through our framework and observe a guiding principle for constructing effective subnetworks: *at a specific sparsity, the top-performing subnetwork always presents a good balance between the number of effective nodes and the number of effective paths.* Inspired by this observation, we present a novel data-agnostic pruning method by solving a multi-objective optimization problem. By conducting extensive experiments across different architectures and datasets, our results demonstrate that our approach outperforms state-of-the-art PaI methods while it is able to discover subnetworks that have much lower inference FLOPs (up to 3.4$\times$). Code will be fully released.

Data pruning, which aims to downsize a large training set into a small informative subset, is crucial for reducing the enormous computational costs of modern deep learning. Though large-scale data collections invariably contain annotation noise and numerous robust learning methods have been developed, data pruning for the noise-robust learning scenario has received little attention. With state-of-the-art Re-labeling methods that self-correct erroneous labels while training, it is challenging to identify which subset induces the most accurate re-labeling of erroneous labels in the entire training set. In this paper, we formalize the problem of data pruning with re-labeling. We first show that the likelihood of a training example being correctly re-labeled is proportional to the prediction confidence of its neighborhood in the subset. Therefore, we propose a novel data pruning algorithm, Prune4Rel, that finds a subset maximizing the total neighborhood confidence of all training examples, thereby maximizing the re-labeling accuracy and generalization performance. Extensive experiments on four real and one synthetic noisy datasets show that Prune4Rel outperforms the baselines with Re-labeling models by up to 9.1% as well as those with a standard model by up to 21.6%.

Graph Structure Learning (GSL) has recently garnered considerable attention due to its ability to optimize both the parameters of Graph Neural Networks (GNNs) and the computation graph structure simultaneously. Despite the proliferation of GSL methods developed in recent years, there is no standard experimental setting or fair comparison for performance evaluation, which creates a great obstacle to understanding the progress in this field. To fill this gap, we systematically analyze the performance of GSL in different scenarios and develop a comprehensive Graph Structure Learning Benchmark (GSLB) curated from 20 diverse graph datasets and 16 distinct GSL algorithms. Specifically, GSLB systematically investigates the characteristics of GSL in terms of three dimensions: effectiveness, robustness, and complexity. We comprehensively evaluate state-of-the-art GSL algorithms in node- and graph-level tasks, and analyze their performance in robust learning and model complexity. Further, to facilitate reproducible research, we have developed an easy-to-use library for training, evaluating, and visualizing different GSL methods. Empirical results of our extensive experiments demonstrate the ability of GSL and reveal its potential benefits on various downstream tasks, offering insights and opportunities for future research. The code of GSLB is available at: https://github.com/GSL-Benchmark/GSLB.

We present a new high-quality validated image caption rating (VICR) dataset. How well a caption fits an image can be difficult to assess due to the subjective nature of caption quality. How do we evaluate whether a caption is good? We generated a new dataset to help answer this question by using our new image caption rating system, which consists of a novel robust rating scale and gamified approach to gathering human ratings. We show that our approach is consistent and teachable. 113 participants were involved in generating the dataset, which is composed of 68,217 ratings among 15,646 image-caption pairs. Our new dataset has greater inter-rater agreement than the state of the art, and custom machine learning rating predictors that were trained on our dataset outperform previous metrics. We improve over Flickr8k-Expert in Kendall's $W$ by 12\% and in Fleiss' $\kappa$ by 19\%, and thus provide a new benchmark dataset for image caption rating.

We investigate the mathematical capabilities of two versions of ChatGPT (released 9-January-2023 and 30-January-2023) and of GPT-4 by testing them on publicly available datasets, as well as hand-crafted ones, using a novel evaluation scheme. In contrast to formal mathematics, where large databases of formal proofs are available (e.g., mathlib, the Lean Mathematical Library), current datasets of natural-language mathematics used to benchmark language models either cover only elementary mathematics or are very small. We address this by publicly releasing two new datasets: GHOSTS and miniGHOSTS. These are the first natural-language datasets curated by working researchers in mathematics that (1) aim to cover graduate-level mathematics, (2) provide a holistic overview of the mathematical capabilities of language models, and (3) distinguish multiple dimensions of mathematical reasoning. These datasets test, by using 1636 human expert evaluations, whether ChatGPT and GPT-4 can be helpful assistants to professional mathematicians by emulating use cases that arise in the daily professional activities of mathematicians. We benchmark the models on a range of fine-grained performance metrics. For advanced mathematics, this is the most detailed evaluation effort to date. We find that ChatGPT and GPT-4 can be used most successfully as mathematical assistants for querying facts, acting as mathematical search engines and knowledge base interfaces. GPT-4 can additionally be used for undergraduate-level mathematics but fails on graduate-level difficulty. Contrary to many positive reports in the media about GPT-4 and ChatGPT's exam-solving abilities (a potential case of selection bias), their overall mathematical performance is well below the level of a graduate student. Hence, if you aim to use ChatGPT to pass a graduate-level math exam, you would be better off copying from your average peer!

We propose DyGFormer, a new Transformer-based architecture for dynamic graph learning. DyGFormer is conceptually simple and only needs to learn from nodes' historical first-hop interactions by: (1) a neighbor co-occurrence encoding scheme that explores the correlations of the source node and destination node based on their historical sequences; (2) a patching technique that divides each sequence into multiple patches and feeds them to Transformer, allowing the model to effectively and efficiently benefit from longer histories. We also introduce DyGLib, a unified library with standard training pipelines, extensible coding interfaces, and comprehensive evaluating protocols to promote reproducible, scalable, and credible dynamic graph learning research. By performing exhaustive experiments on thirteen datasets for dynamic link prediction and dynamic node classification tasks, we find that DyGFormer achieves state-of-the-art performance on most of the datasets, demonstrating its effectiveness in capturing nodes' correlations and long-term temporal dependencies. Moreover, some results of baselines are inconsistent with previous reports, which may be caused by their diverse but less rigorous implementations, showing the importance of DyGLib. All the used resources are publicly available at https://github.com/yule-BUAA/DyGLib.

Spectral clustering has gained popularity for clustering non-convex data due to its simplicity and effectiveness. It is essential to construct a similarity graph using a high-quality affinity measure that models the local neighborhood relations among the data samples. However, incomplete data can lead to inaccurate affinity measures, resulting in degraded clustering performance. To address these issues, we propose an imputation-free framework with two novel approaches to improve spectral clustering on incomplete data. Firstly, we introduce a new kernel correction method that enhances the quality of the kernel matrix estimated on incomplete data with a theoretical guarantee, benefiting classical spectral clustering on pre-defined kernels. Secondly, we develop a series of affinity learning methods that equip the self-expressive framework with $\ell_p$-norm to construct an intrinsic affinity matrix with an adaptive extension. Our methods outperform existing data imputation and distance calibration techniques on benchmark datasets, offering a promising solution to spectral clustering on incomplete data in various real-world applications.

This paper proposes a novel strategy for estimating the heterogeneous treatment effect called the Fused and Accurate Shrinkage Tree ($\mathrm{FAST}$). Our approach utilizes both trial and observational data to improve the accuracy and robustness of the estimator. Inspired by the concept of shrinkage estimation in statistics, we develop an optimal weighting scheme and a corresponding estimator that balances the unbiased estimator based on the trial data with the potentially biased estimator based on the observational data. Specifically, combined with tree-based techniques, we introduce a new split criterion that utilizes both trial data and observational data to more accurately estimate the treatment effect. Furthermore, we confirm the consistency of our proposed tree-based estimator and demonstrate the effectiveness of our criterion in reducing prediction error through theoretical analysis. The advantageous finite sample performance of the $\mathrm{FAST}$ and its ensemble version over existing methods is demonstrated via simulations and real data analysis.

Code completion models have made significant progress in recent years, yet current popular evaluation datasets, such as HumanEval and MBPP, predominantly focus on code completion tasks within a single file. This over-simplified setting falls short of representing the real-world software development scenario where repositories span multiple files with numerous cross-file dependencies, and accessing and understanding cross-file context is often required to complete the code correctly. To fill in this gap, we propose CrossCodeEval, a diverse and multilingual code completion benchmark that necessitates an in-depth cross-file contextual understanding to complete the code accurately. CrossCodeEval is built on a diverse set of real-world, open-sourced, permissively-licensed repositories in four popular programming languages: Python, Java, TypeScript, and C#. To create examples that strictly require cross-file context for accurate completion, we propose a straightforward yet efficient static-analysis-based approach to pinpoint the use of cross-file context within the current file. Extensive experiments on state-of-the-art code language models like CodeGen and StarCoder demonstrate that CrossCodeEval is extremely challenging when the relevant cross-file context is absent, and we see clear improvements when adding these context into the prompt. However, despite such improvements, the pinnacle of performance remains notably unattained even with the highest-performing model, indicating that CrossCodeEval is also capable of assessing model's capability in leveraging extensive context to make better code completion. Finally, we benchmarked various methods in retrieving cross-file context, and show that CrossCodeEval can also be used to measure the capability of code retrievers.

This paper introduces an efficient and robust framework for hybrid query (HQ) processing, which combines approximate nearest neighbor search (ANNS) with attribute constraint. HQ aims to find objects that are similar to a feature vector and match some structured attributes. Existing methods handle ANNS and attribute filtering separately, leading to inefficiency and inaccuracy. Our framework, called native hybrid query (NHQ), builds a composite index based on proximity graph (PG) and applies joint pruning for HQ. We can easily adapt existing PGs to this framework for efficient HQ processing. We also propose two new navigable PGs (NPGs) with optimized edge selection and routing, which improve the overall ANNS performance. We implement five HQ methods based on the proposed NPGs and existing PGs in NHQ, and show that they outperform the state-of-the-art methods on 10 real-world datasets (up to 315$\times$ faster with the same accuracy).

We present the first diffusion-based framework that can learn an unknown distribution using only highly-corrupted samples. This problem arises in scientific applications where access to uncorrupted samples is impossible or expensive to acquire. Another benefit of our approach is the ability to train generative models that are less likely to memorize any individual training sample, since they never observe clean training data. Our main idea is to introduce additional measurement distortion during the diffusion process and require the model to predict the original corrupted image from the further corrupted image. We prove that our method leads to models that learn the conditional expectation of the full uncorrupted image given this additional measurement corruption. This holds for any corruption process that satisfies some technical conditions (and in particular includes inpainting and compressed sensing). We train models on standard benchmarks (CelebA, CIFAR-10 and AFHQ) and show that we can learn the distribution even when all the training samples have 90\% of their pixels missing. We also show that we can finetune foundation models on small corrupted datasets (e.g. MRI scans with block corruptions) and learn the clean distribution without memorizing the training set.

Naturalistic stimuli evoke complex neural responses with spatial and temporal properties that differ across individuals. Current alignment methods focus on either spatial hyperalignment (assuming exact temporal correspondence) or temporal alignment (assuming exact spatial correspondence). Here, we propose a hybrid model, the Hyper-HMM, that simultaneously aligns both temporal and spatial features across brains. The model learns to linearly project voxels to a reduced-dimension latent space, in which timecourses are segmented into corresponding temporal events. This approach allows tracking of each individual's mental trajectory through an event sequence, and also allows for alignment with other feature spaces such as stimulus content. Using an fMRI dataset in which students watch videos of class lectures, we demonstrate that the Hyper-HMM can be used to map all participants and the semantic content of the videos into a common low-dimensional space, and that these mappings generalize to held-out data. Our model provides a new window into individual cognitive dynamics evoked by complex naturalistic stimuli.

In this work we explore recent advances in instruction-tuning language models on a range of open instruction-following datasets. Despite recent claims that open models can be on par with state-of-the-art proprietary models, these claims are often accompanied by limited evaluation, making it difficult to compare models across the board and determine the utility of various resources. We provide a large set of instruction-tuned models from 6.7B to 65B parameters in size, trained on 12 instruction datasets ranging from manually curated (e.g., OpenAssistant) to synthetic and distilled (e.g., Alpaca) and systematically evaluate them on their factual knowledge, reasoning, multilinguality, coding, safety, and open-ended instruction following abilities through a collection of automatic, model-based, and human-based metrics. We further introduce Tülu, our best performing instruction-tuned model suite finetuned on a combination of high-quality open resources. Our experiments show that different instruction-tuning datasets can uncover or enhance specific skills, while no single dataset (or combination) provides the best performance across all evaluations. Interestingly, we find that model and human preference-based evaluations fail to reflect differences in model capabilities exposed by benchmark-based evaluations, suggesting the need for the type of systemic evaluation performed in this work. Our evaluations show that the best model in any given evaluation reaches on average 87% of ChatGPT performance, and 73% of GPT-4 performance, suggesting that further investment in building better base models and instruction-tuning data is required to close the gap. We release our instruction-tuned models, including a fully finetuned 65B Tülu, along with our code, data, and evaluation framework to facilitate future research.

Long-tailed object detection (LTOD) aims to handle the extreme data imbalance in real-world datasets, where many tail classes have scarce instances. One popular strategy is to explore extra data with image-level labels, yet it produces limited results due to (1) semantic ambiguity---an image-level label only captures a salient part of the image, ignoring the remaining rich semantics within the image; and (2) location sensitivity---the label highly depends on the locations and crops of the original image, which may change after data transformations like random cropping. To remedy this, we propose RichSem, a simple but effective method, which is robust to learn rich semantics from coarse locations without the need of accurate bounding boxes. RichSem leverages rich semantics from images, which are then served as additional ``soft supervision'' for training detectors. Specifically, we add a semantic branch to our detector to learn these soft semantics and enhance feature representations for long-tailed object detection. The semantic branch is only used for training and is removed during inference. RichSem achieves consistent improvements on both overall and rare-category of LVIS under different backbones and detectors. Our method achieves state-of-the-art performance without requiring complex training and testing procedures. Moreover, we show the effectiveness of our method on other long-tailed datasets with additional experiments.

We introduce HT-Step, a large-scale dataset containing temporal annotations of instructional article steps in cooking videos. It includes 122k segment-level annotations over 20k narrated videos (approximately 2.3k hours) of the HowTo100M dataset. Each annotation provides a temporal interval, and a categorical step label from a taxonomy of 4,958 unique steps automatically mined from wikiHow articles which include rich descriptions of each step. Our dataset significantly surpasses existing labeled step datasets in terms of scale, number of tasks, and richness of natural language step descriptions. Based on these annotations, we introduce a strongly supervised benchmark for aligning instructional articles with how-to videos and present a comprehensive evaluation of baseline methods for this task. By publicly releasing these annotations and defining rigorous evaluation protocols and metrics, we hope to significantly accelerate research in the field of procedural activity understanding.