NeurIPS 2024 论文列表

Achieving robust 3D perception in the face of corrupted data presents an challenging hurdle within 3D vision research. Contemporary transformer-based point cloud recognition models, albeit advanced, tend to overfit to specific patterns, consequently undermining their robustness against corruption. In this work, we introduce the Target-Guided Adversarial Point Cloud Transformer, termed APCT, a novel architecture designed to augment global structure capture through an adversarial feature erasing mechanism predicated on patterns discerned at each step during training. Specifically, APCT integrates an Adversarial Significance Identifier and a Target-guided Promptor. The Adversarial Significance Identifier, is tasked with discerning token significance by integrating global contextual analysis, utilizing a structural salience index algorithm alongside an auxiliary supervisory mechanism. The Target-guided Promptor, is responsible for accentuating the propensity for token discard within the self-attention mechanism, utilizing the value derived above, consequently directing the model attention towards alternative segments in subsequent stages. By iteratively applying this strategy in multiple steps during training, the network progressively identifies and integrates an expanded array of object-associated patterns. Extensive experiments demonstrate that our method achieves state-of-the-art results on multiple corruption benchmarks.

This paper investigates an under-explored challenge in large language models (LLMs): chain-of-thought prompting with noisy rationales, which include irrelevant or inaccurate reasoning thoughts within examples used for in-context learning. We construct NoRa dataset that is tailored to evaluate the robustness of reasoning in the presence of noisy rationales. Our findings on NoRa dataset reveal a prevalent vulnerability to such noise among current LLMs, with existing robust methods like self-correction and self-consistency showing limited efficacy. Notably, compared to prompting with clean rationales, base LLM drops by 1.4%-19.8% in accuracy with irrelevant thoughts and more drastically by 2.2%-40.4% with inaccurate thoughts. Addressing this challenge necessitates external supervision that should be accessible in practice. Here, we propose the method of contrastive denoising with noisy chain-of-thought (CD-CoT). It enhances LLMs' denoising-reasoning capabilities by contrasting noisy rationales with only one clean rationale, which can be the minimal requirement for denoising-purpose prompting. This method follows a principle of exploration and exploitation: (1) rephrasing and selecting rationales in the input space to achieve explicit denoising and (2) exploring diverse reasoning paths and voting on answers in the output space. Empirically, CD-CoT demonstrates an average improvement of 17.8% in accuracy over the base model and shows significantly stronger denoising capabilities than baseline methods. The source code is publicly available at: https://github.com/tmlr-group/NoisyRationales.

In-context learning is an emergent paradigm in large language models (LLMs) that enables them to generalize to new tasks and domains by simply prompting these models with a few exemplars without explicit parameter updates. Many attempts have been made to understand in-context learning in LLMs as a function of model scale, pretraining data, and other factors. In this work, we propose a new mechanism to probe and understand in-context learning from the lens of decision boundaries for in-context binary classification. Decision boundaries are straightforward to visualize and provide important information about the qualitative behavior of the inductive biases of standard classifiers. To our surprise, we find that the decision boundaries learned by current LLMs in simple binary classification tasks are often irregularly non-smooth, regardless of task linearity. This paper investigates the factors influencing these decision boundaries and explores methods to enhance their generalizability. We assess various approaches, including training-free and fine-tuning methods for LLMs, the impact of model architecture, and the effectiveness of active prompting techniques for smoothing decision boundaries in a data-efficient manner. Our findings provide a deeper understanding of in-context learning dynamics and offer practical improvements for enhancing robustness and generalizability of in-context learning.

Availability attacks provide a tool to prevent the unauthorized use of private data and commercial datasets by generating imperceptible noise and crafting unlearnable examples before release. Ideally, the obtained unlearnability can prevent algorithms from training usable models. When supervised learning (SL) algorithms have failed, a malicious data collector possibly resorts to contrastive learning (CL) algorithms to bypass the protection. Through evaluation, we have found that most existing methods are unable to achieve both supervised and contrastive unlearnability, which poses risks to data protection by availability attacks. Different from recent methods based on contrastive learning, we employ contrastive-like data augmentations in supervised learning frameworks to obtain attacks effective for both SL and CL. Our proposed AUE and AAP attacks achieve state-of-the-art worst-case unlearnability across SL and CL algorithms with less computation consumption, showcasing prospects in real-world applications. The code is available at https://github.com/EhanW/AUE-AAP.

Robotics, autonomous driving, augmented reality, and many embodied computer vision applications must quickly react to user-defined events unfolding in real time. We address this setting by proposing a novel task for multimodal video understanding---Streaming Detection of Queried Event Start (SDQES). The goal of SDQES is to identify the beginning of a complex event as described by a natural language query, with high accuracy and low latency. We introduce a new benchmark based on the Ego4D dataset, as well as new task-specific metrics to study streaming multimodal detection of diverse events in an egocentric video setting. Inspired by parameter-efficient fine-tuning methods in NLP and for video tasks, we propose adapter-based baselines that enable image-to-video transfer learning, allowing for efficient online video modeling. We evaluate three vision-language backbones and three adapter architectures on both short-clip and untrimmed video settings.

Motivated by the problem of compressing point sets into as few bits as possible while maintaining information about approximate distances between points, we construct random nonlinear maps $\varphi_\ell$ that compress point sets in the following way. For a point set $S$, the map $\varphi_\ell:\mathbb{R}^d \to N^{-1/2}\{-1,1\}^N$ has the property that storing $\varphi_\ell(S)$ (a sketch of $S$) allows one to report squared distances between points up to some multiplicative $(1\pm \epsilon)$ error with high probability. The maps $\varphi_\ell$ are the $\ell$-fold composition of a certain type of random feature mapping. Compared to existing techniques, our maps offer several advantages. The standard method for compressing point sets by random mappings relies on the Johnson-Lindenstrauss lemma and involves compressing point sets with a random linear map. The main advantage of our maps $\varphi_\ell$ over random linear maps is that ours map point sets directly into the discrete cube $N^{-1/2}\{-1,1\}^N$ and so there is no additional step needed to convert the sketch to bits. For some range of parameters, our maps $\varphi_\ell$ produce sketches using fewer bits of storage space. We validate the method with experiments, including an application to nearest neighbor search.

Videos carry rich visual information including object description, action, interaction, etc., but the existing multimodal large language models (MLLMs) fell short in referential understanding scenarios such as video-based referring. In this paper, we present Artemis, an MLLM that pushes video-based referential understanding to a finer level. Given a video, Artemis receives a natural-language question with a bounding box in any video frame and describes the referred target in the entire video. The key to achieving this goal lies in extracting compact, target-specific video features, where we set a solid baseline by tracking and selecting spatiotemporal features from the video. We train Artemis on the newly established ViderRef45K dataset with 45K video-QA pairs and design a computationally efficient, three-stage training procedure. Results are promising both quantitatively and qualitatively. Additionally, we show that Artemis can be integrated with video grounding and text summarization tools to understand more complex scenarios. Code and data are available at https://github.com/NeurIPS24Artemis/Artemis.

Large vision language models (LVLMs) integrate large language models (LLMs) with pre-trained vision encoders, thereby activating the perception capability of the model to understand image inputs for different queries and conduct subsequent reasoning. Improving this capability requires high-quality vision-language data, which is costly and labor-intensive to acquire. Self-training approaches have been effective in single-modal settings to alleviate the need for labeled data by leveraging model's own generation. However, effective self-training remains a challenge regarding the unique visual perception and reasoning capability of LVLMs. To address this, we introduce **S**elf-**T**raining on **I**mage **C**omprehension (**STIC**), which emphasizes a self-training approach specifically for image comprehension. First, the model self-constructs a preference dataset for image descriptions using unlabeled images. Preferred responses are generated through a step-by-step prompt, while dis-preferred responses are generated from either corrupted images or misleading prompts. To further self-improve reasoning on the extracted visual information, we let the model reuse a small portion of existing instruction-tuning data and append its self-generated image descriptions to the prompts. We validate the effectiveness of STIC across seven different benchmarks, demonstrating substantial performance gains of 4.0% on average while using 70% less supervised fine-tuning data than the current method. Further studies dive into various components of STIC and highlight its potential to leverage vast quantities of unlabeled images for self-training.

Learning from AI feedback (LAIF) is a popular paradigm for improving the instruction-following abilities of powerful pre-trained language models. LAIF first performs supervised fine-tuning (SFT) using demonstrations from a teacher model and then further fine-tunes the model with reinforcement learning (RL) or direct preference optimization (DPO), using feedback from a critic model. While recent popular open-source models have demonstrated substantial improvements in performance from the RL step, in this paper we question whether the complexity of this RL step is truly warranted for AI feedback. We show that the improvements of the RL step are virtually entirely due to the widespread practice of using a weaker teacher model (e.g. GPT-3.5) for SFT data collection than the critic (e.g., GPT-4) used for AI feedback generation. Specifically, we show that simple supervised fine-tuning with GPT-4 as the teacher outperforms existing LAIF pipelines. More generally, we find that the gains from LAIF vary substantially across base model families, test-time evaluation protocols, and critic models. Finally, we provide a mechanistic explanation for when SFT may outperform the full two-step LAIF pipeline as well as suggestions for making LAIF maximally useful in practice.

Motivated by the problem of detecting AI-generated text, we consider the problem of watermarking the output of language models with provable guarantees. We aim for watermarks which satisfy: (a) undetectability, a cryptographic notion introduced by Christ, Gunn, & Zamir (2023) which stipulates that it is computationally hard to distinguish watermarked language model outputs from the model's actual output distribution; and (b) robustness to channels which introduce a constant fraction of adversarial insertions, substitutions, and deletions to the watermarked text. Earlier schemes could only handle stochastic substitutions and deletions, and thus we are aiming for a more natural and appealing robustness guarantee that holds with respect to edit distance. Our main result is a watermarking scheme which achieves both (a) and (b) when the alphabet size for the language model is allowed to grow as a polynomial in the security parameter. To derive such a scheme, we follow an approach introduced by Christ & Gunn (2024), which proceeds via first constructing pseudorandom codes satisfying undetectability and robustness properties analogous to those above; our codes have the additional benefit of relying on weaker computational assumptions than used in previous work. Then we show that there is a generic transformation from such codes over large alphabets to watermarking schemes for arbitrary language models.

The CLIP network excels in various tasks, but struggles with text-visual images i.e., images that contain both text and visual objects; it risks confusing textual and visual representations. To address this issue, we propose MirrorCLIP, a zero-shot framework, which disentangles the image features of CLIP by exploiting the difference in the mirror effect between visual objects and text in the images. Specifically, MirrorCLIP takes both original and flipped images as inputs, comparing their features dimension-wise in the latent space to generate disentangling masks. With disentangling masks, we further design filters to separate textual and visual factors more precisely, and then get disentangled representations. Qualitative experiments using stable diffusion models and class activation mapping (CAM) validate the effectiveness of our disentanglement. Moreover, our proposed MirrorCLIP reduces confusion when encountering text-visual images and achieves a substantial improvement on typographic defense, further demonstrating its superior ability of disentanglement. Our code is available at https://github.com/tcwangbuaa/MirrorCLIP

Prior multi-frame optical flow methods typically estimate flow repeatedly in a pair-wise manner, leading to significant computational redundancy. To mitigate this, we implement a Streamlined In-batch Multi-frame (SIM) pipeline, specifically tailored to video inputs to minimize redundant calculations. It enables the simultaneous prediction of successive unidirectional flows in a single forward pass, boosting processing speed by 44.43% and reaching efficiencies on par with two-frame networks. Moreover, we investigate various spatiotemporal modeling methods for optical flow estimation within this pipeline. Notably, we propose a simple yet highly effective parameter-efficient Integrative spatiotemporal Coherence (ISC) modeling method, alongside a lightweight Global Temporal Regressor (GTR) to harness temporal cues. The proposed ISC and GTR bring powerful spatiotemporal modeling capabilities and significantly enhance accuracy, including in occluded areas, while adding modest computations to the SIM pipeline. Compared to the baseline, our approach, StreamFlow, achieves performance enhancements of 15.45% and 11.37% on the Sintel clean and final test sets respectively, with gains of 15.53% and 10.77% on occluded regions and only a 1.11% rise in latency. Furthermore, StreamFlow exhibits state-of-the-art cross-dataset testing results on Sintel and KITTI, demonstrating its robust cross-domain generalization capabilities. The code is available [here](https://github.com/littlespray/StreamFlow).

The Neural Tangent Kernel (NTK) viewpoint is widely employed to analyze the training dynamics of overparameterized Physics-Informed Neural Networks (PINNs). However, unlike the case of linear Partial Differential Equations (PDEs), we show how the NTK perspective falls short in the nonlinear scenario. Specifically, we establish that the NTK yields a random matrix at initialization that is not constant during training, contrary to conventional belief. Another significant difference from the linear regime is that, even in the idealistic infinite-width limit, the Hessian does not vanish and hence it cannot be disregarded during training. This motivates the adoption of second-order optimization methods. We explore the convergence guarantees of such methods in both linear and nonlinear cases, addressing challenges such as spectral bias and slow convergence. Every theoretical result is supported by numerical examples with both linear and nonlinear PDEs, and we highlight the benefits of second-order methods in benchmark test cases.

Class-incremental learning (CIL) aims to continually learn a sequence of tasks, with each task consisting of a set of unique classes. Graph CIL (GCIL) follows the same setting but needs to deal with graph tasks (e.g., node classification in a graph). The key characteristic of CIL lies in the absence of task identifiers (IDs) during inference, which causes a significant challenge in separating classes from different tasks (i.e., inter-task class separation). Being able to accurately predict the task IDs can help address this issue, but it is a challenging problem. In this paper, we show theoretically that accurate task ID prediction on graph data can be achieved by a Laplacian smoothing-based graph task profiling approach, in which each graph task is modeled by a task prototype based on Laplacian smoothing over the graph. It guarantees that the task prototypes of the same graph task are nearly the same with a large smoothing step, while those of different tasks are distinct due to differences in graph structure and node attributes. Further, to avoid the catastrophic forgetting of the knowledge learned in previous graph tasks, we propose a novel graph prompting approach for GCIL which learns a small discriminative graph prompt for each task, essentially resulting in a separate classification model for each task. The prompt learning requires the training of a single graph neural network (GNN) only once on the first task, and no data replay is required thereafter, thereby obtaining a GCIL model being both replay-free and forget-free. Extensive experiments on four GCIL benchmarks show that i) our task prototype-based method can achieve 100% task ID prediction accuracy on all four datasets, ii) our GCIL model significantly outperforms state-of-the-art competing methods by at least 18% in average CIL accuracy, and iii) our model is fully free of forgetting on the four datasets.

Benchmarking vision-based driving policies is challenging. On one hand, open-loop evaluation with real data is easy, but these results do not reflect closed-loop performance. On the other, closed-loop evaluation is possible in simulation, but is hard to scale due to its significant computational demands. Further, the simulators available today exhibit a large domain gap to real data. This has resulted in an inability to draw clear conclusions from the rapidly growing body of research on end-to-end autonomous driving. In this paper, we present NAVSIM, a middle ground between these evaluation paradigms, where we use large datasets in combination with a non-reactive simulator to enable large-scale real-world benchmarking. Specifically, we gather simulation-based metrics, such as progress and time to collision, by unrolling bird's eye view abstractions of the test scenes for a short simulation horizon. Our simulation is non-reactive, i.e., the evaluated policy and environment do not influence each other. As we demonstrate empirically, this decoupling allows open-loop metric computation while being better aligned with closed-loop evaluations than traditional displacement errors. NAVSIM enabled a new competition held at CVPR 2024, where 143 teams submitted 463 entries, resulting in several new insights. On a large set of challenging scenarios, we observe that simple methods with moderate compute requirements such as TransFuser can match recent large-scale end-to-end driving architectures such as UniAD. Our modular framework can potentially be extended with new datasets, data curation strategies, and metrics, and will be continually maintained to host future challenges. Our code is available at https://github.com/autonomousvision/navsim.

LLMs are computationally expensive to pre-train due to their large scale. Model growth emerges as a promising approach by leveraging smaller models to accelerate the training of larger ones. However, the viability of these model growth methods in efficient LLM pre-training remains underexplored. This work identifies three critical $\underline{\textit{O}}$bstacles: ($\textit{O}$1) lack of comprehensive evaluation, ($\textit{O}$2) untested viability for scaling, and ($\textit{O}$3) lack of empirical guidelines. To tackle $\textit{O}$1, we summarize existing approaches into four atomic growth operators and systematically evaluate them in a standardized LLM pre-training setting. Our findings reveal that a depthwise stacking operator, called $G_{\text{stack}}$, exhibits remarkable acceleration in training, leading to decreased loss and improved overall performance on eight standard NLP benchmarks compared to strong baselines. Motivated by these promising results, we conduct extensive experiments to delve deeper into $G_{\text{stack}}$ to address $\textit{O}$2 and $\textit{O}$3. For $\textit{O}$2 (untested scalability), our study shows that $G_{\text{stack}}$ is scalable and consistently performs well, with experiments up to 7B LLMs after growth and pre-training LLMs with 750B tokens. For example, compared to a conventionally trained 7B model using 300B tokens, our $G_{\text{stack}}$ model converges to the same loss with 194B tokens, resulting in a 54.6\% speedup. We further address $\textit{O}$3 (lack of empirical guidelines) by formalizing guidelines to determine growth timing and growth factor for $G_{\text{stack}}$, making it practical in general LLM pre-training. We also provide in-depth discussions and comprehensive ablation studies of $G_{\text{stack}}$. Our code and pre-trained model are available at https://llm-stacking.github.io/.

We introduce Language Feedback Models (LFMs) that identify desirable behaviour --- actions that help achieve tasks specified in the instruction - for imitation learning in instruction following. To train LFMs, we obtain feedback from Large Language Models (LLMs) on visual trajectories verbalized to language descriptions. First, by using LFMs to identify desirable behaviour to imitate, we improve in task-completion rate over strong behavioural cloning baselines on three distinct language grounding environments (Touchdown, ScienceWorld, and ALFWorld). Second, LFMs outperform using LLMs as experts to directly predict actions, when controlling for the number of LLM output tokens. Third, LFMs generalize to unseen environments, improving task-completion rate by 3.5-12.0% through one round of adaptation. Finally, LFMs can be modified to provide human-interpretable feedback without performance loss, allowing human verification of desirable behaviour for imitation learning.

Generative diffusion models and many stochastic models in science and engineering naturally live in infinite dimensions before discretisation. To incorporate observed data for statistical and learning tasks, one needs to condition on observations. While recent work has treated conditioning linear processes in infinite dimensions, conditioning non-linear processes in infinite dimensions has not been explored. This paper conditions function valued stochastic processes without prior discretisation. To do so, we use an infinite-dimensional version of Girsanov's theorem to condition a function-valued stochastic process, leading to a stochastic differential equation (SDE) for the conditioned process involving the score. We apply this technique to do time series analysis for shapes of organisms in evolutionary biology, where we discretise via the Fourier basis and then learn the coefficients of the score function with score matching methods.

Pre-trained language encoders---functions that represent text as vectors---are an integral component of many NLP tasks. We tackle a natural question in language encoder analysis: What does it mean for two encoders to be similar? We contend that a faithful measure of similarity needs to be \emph{intrinsic}, that is, task-independent, yet still be informative of \emph{extrinsic} similarity---the performance on downstream tasks. It is common to consider two encoders similar if they are \emph{homotopic}, i.e., if they can be aligned through some transformation. In this spirit, we study the properties of \emph{affine} alignment of language encoders and its implications on extrinsic similarity. We find that while affine alignment is fundamentally an asymmetric notion of similarity, it is still informative of extrinsic similarity. We confirm this on datasets of natural language representations. Beyond providing useful bounds on extrinsic similarity, affine intrinsic similarity also allows us to begin uncovering the structure of the space of pre-trained encoders by defining an order over them.