NeurIPS 2024 论文列表

The essence of multi-modal fusion lies in exploiting the complementary information inherent in diverse modalities.However, most prevalent fusion methods rely on traditional neural architectures and are inadequately equipped to capture the dynamics of interactions across modalities, particularly in presence of complex intra- and inter-modality correlations.Recent advancements in State Space Models (SSMs), notably exemplified by the Mamba model, have emerged as promising contenders. Particularly, its state evolving process implies stronger modality fusion paradigm, making multi-modal fusion on SSMs an appealing direction. However, fusing multiple modalities is challenging for SSMs due to its hardware-aware parallelism designs. To this end, this paper proposes the Coupled SSM model, for coupling state chains of multiple modalities while maintaining independence of intra-modality state processes. Specifically, in our coupled scheme, we devise an inter-modal hidden states transition scheme, in which the current state is dependent on the states of its own chain and that of the neighbouring chains at the previous time-step. To fully comply with the hardware-aware parallelism, we obtain the global convolution kernel by deriving the state equation while introducing the historical state.Extensive experiments on CMU-MOSEI, CH-SIMS, CH-SIMSV2 through multi-domain input verify the effectiveness of our model compared to current state-of-the-art methods, improved F1-Score by 0.4%, 0.9%, and 2.3% on the three datasets respectively, 49% faster inference and 83.7% GPU memory save. The results demonstrate that Coupled Mamba model is capable of enhanced multi-modal fusion.

Recent advancements in State Space Models, notably Mamba, have demonstrated superior performance over the dominant Transformer models, particularly in reducing the computational complexity from quadratic to linear. Yet, difficulties in adapting Mamba from language to vision tasks arise due to the distinct characteristics of visual data, such as the spatial locality and adjacency within images and large variations in information granularity across visual tokens. Existing vision Mamba approaches either flatten tokens into sequences in a raster scan fashion, which breaks the local adjacency of images, or manually partition tokens into windows, which limits their long-range modeling and generalization capabilities. To address these limitations, we present a new vision Mamba model, coined QuadMamba, that effectively captures local dependencies of varying granularities via quadtree-based image partition and scan. Concretely, our lightweight quadtree-based scan module learns to preserve the 2D locality of spatial regions within learned window quadrants. The module estimates the locality score of each token from their features, before adaptively partitioning tokens into window quadrants. An omnidirectional window shifting scheme is also introduced to capture more intact and informative features across different local regions. To make the discretized quadtree partition end-to-end trainable, we further devise a sequence masking strategy based on Gumbel-Softmax and its straight-through gradient estimator. Extensive experiments demonstrate that QuadMamba achieves state-of-the-art performance in various vision tasks, including image classification, object detection, instance segmentation, and semantic segmentation. Our code and models will be released.

When machine learning systems face dataset shift, model calibration plays a pivotal role in ensuring their reliability. Calibration error (CE) provides insights into the alignment between the predicted confidence scores and the classifier accuracy. While prior works have delved into the implications of dataset shift on calibration, existing CE estimators either (i) assume access to labeled data from the target domain, often unavailable in practice, or (ii) are derived under a covariate shift assumption. In this work we propose a novel, label-free, consistent CE estimator under label shift. Label shift is characterized by changes in the marginal label distribution p(Y), with a constant conditional p(X|Y) distribution between the source and target. We introduce a novel calibration method, called LaSCal, which uses the estimator in conjunction with a post-hoc calibration strategy, to perform unsupervised calibration on the target distribution. Our thorough empirical analysis demonstrates the effectiveness and reliability of the proposed approach across different modalities, model architectures and label shift intensities.

Time series forecasting typically needs to address non-stationary data with evolving trend and seasonal patterns. To address the non-stationarity, reversible instance normalization has been recently proposed to alleviate impacts from the trend with certain statistical measures, e.g., mean and variance. Although they demonstrate improved predictive accuracy, they are limited to expressing basic trends and are incapable of handling seasonal patterns. To address this limitation, this paper proposes a new instance normalization solution, called frequency adaptive normalization (FAN), which extends instance normalization in handling both dynamic trend and seasonal patterns. Specifically, we employ the Fourier transform to identify instance-wise predominant frequent components that cover most non-stationary factors. Furthermore, the discrepancy of those frequency components between inputs and outputs is explicitly modeled as a prediction task with a simple MLP model. FAN is a model-agnostic method that can be applied to arbitrary predictive backbones. We instantiate FAN on four widely used forecasting models as the backbone and evaluate their prediction performance improvements on eight benchmark datasets. FAN demonstrates significant performance advancement, achieving 7.76\%$\sim$37.90\% average improvements in MSE. Our code is publicly available at http://github.com/icannotnamemyself/FAN.

Despite satisfactory results on ``easy'' cases of single image reflection separation, prior dual-stream methods still suffer from considerable performance degradation when facing complex ones, i.e, the transmission layer is densely entangled with the reflection having a wide distribution of spatial intensity. The main reasons come from the lack of concern on the feature correlation during interaction, and the limited receptive field. To remedy these deficiencies, this paper presents a Dual-Stream Interactive Transformer (DSIT) design. Specifically, we devise a dual-attention interactive structure that embraces a dual-stream self-attention and a layer-aware dual-stream cross-attention mechanism to simultaneously capture intra-layer and inter-layer feature correlations. Meanwhile, the introduction of attention mechanisms can also mitigate the receptive field limitation. We modulate single-stream pre-trained Transformer embeddings with dual-stream convolutional features through cross-architecture interactions to provide richer semantic priors, thereby further relieving the ill-posedness of the problem. Extensive experimental results reveal the merits of the proposed DSIT over other state-of-the-art alternatives. Our code is publicly available at https://github.com/mingcv/DSIT.

With the rapid development of multimedia technology, audio-visual learning has emerged as a promising research topic within the field of multimodal analysis. In this paper, we explore parameter-efficient transfer learning for audio-visual learning and propose the Audio-Visual Mixture of Experts (\ourmethodname) to inject adapters into pre-trained models flexibly. Specifically, we introduce unimodal and cross-modal adapters as multiple experts to specialize in intra-modal and inter-modal information, respectively, and employ a lightweight router to dynamically allocate the weights of each expert according to the specific demands of each task. Extensive experiments demonstrate that our proposed approach \ourmethodname achieves superior performance across multiple audio-visual tasks, including AVE, AVVP, AVS, and AVQA. Furthermore, visual-only experimental results also indicate that our approach can tackle challenging scenes where modality information is missing. The source code is available at \url{https://github.com/yingchengy/AVMOE}.

We present Parametric Piecewise Linear Networks (PPLNs) for temporal vision inference. Motivated by the neuromorphic principles that regulate biological neural behaviors, PPLNs are ideal for processing data captured by event cameras, which are built to simulate neural activities in the human retina. We discuss how to represent the membrane potential of an artificial neuron by a parametric piecewise linear function with learnable coefficients. This design echoes the idea of building deep models from learnable parametric functions recently popularized by Kolmogorov–Arnold Networks (KANs). Experiments demonstrate the state-of-the-art performance of PPLNs in event-based and image-based vision applications, including steering prediction, human pose estimation, and motion deblurring.

Global convolutions have shown increasing promise as powerful general-purpose sequence models. However, training long convolutions is challenging, and kernel parameterizations must be able to learn long-range dependencies without overfitting. This work introduces reparameterized multi-resolution convolutions ($\texttt{MRConv}$), a novel approach to parameterizing global convolutional kernels for long-sequence modeling. By leveraging multi-resolution convolutions, incorporating structural reparameterization and introducing learnable kernel decay, $\texttt{MRConv}$ learns expressive long-range kernels that perform well across various data modalities. Our experiments demonstrate state-of-the-art performance on the Long Range Arena, Sequential CIFAR, and Speech Commands tasks among convolution models and linear-time transformers. Moreover, we report improved performance on ImageNet classification by replacing 2D convolutions with 1D $\texttt{MRConv}$ layers.

We describe a family of architectures to support transductive inference by allowing memory to grow to a finite but a-priori unknown bound while making efficient use of finite resources for inference. Current architectures use such resources to represent data either eidetically over a finite span ('context' in Transformers), or fading over an infinite span (in State Space Models, or SSMs). Recent hybrid architectures have combined eidetic and fading memory, but with limitations that do not allow the designer or the learning process to seamlessly modulate the two, nor to extend the eidetic memory span. We leverage ideas from Stochastic Realization Theory to develop a class of models called B'MOJO to seamlessly combine eidetic and fading memory within an elementary composable module. The overall architecture can be used to implement models that can access short-term eidetic memory 'in-context,' permanent structural memory 'in-weights,' fading memory 'in-state,' and long-term eidetic memory 'in-storage' by natively incorporating retrieval from an asynchronously updated memory. We show that Transformers, existing SSMs such as Mamba, and hybrid architectures such as Jamba are special cases of B'MOJO and describe a basic implementation that can be stacked and scaled efficiently in hardware. We test B'MOJO on transductive inference tasks, such as associative recall, where it outperforms existing SSMs and Hybrid models; as a baseline, we test ordinary language modeling where B'MOJO achieves perplexity comparable to similarly-sized Transformers and SSMs up to 1.4B parameters, while being up to 10% faster to train. Finally, we test whether models trained inductively on a-priori bounded sequences (up to 8K tokens) can still perform transductive inference on sequences many-fold longer. B'MOJO's ability to modulate eidetic and fading memory results in better inference on longer sequences tested up to 32K tokens, four-fold the length of the longest sequences seen during training.

Widely adopted in modern Vision Transformer designs, Softmax attention can effectively capture long-range visual information; however, it incurs excessive computational cost when dealing with high-resolution inputs. In contrast, linear attention naturally enjoys linear complexity and has great potential to scale up to higher-resolution images. Nonetheless, the unsatisfactory performance of linear attention greatly limits its practical application in various scenarios. In this paper, we take a step forward to close the gap between the linear and Softmax attention with novel theoretical analyses, which demystify the core factors behind the performance deviations. Specifically, we present two key perspectives to understand and alleviate the limitations of linear attention: the injective property and the local modeling ability. Firstly, we prove that linear attention is not injective, which is prone to assign identical attention weights to different query vectors, thus adding to severe semantic confusion since different queries correspond to the same outputs. Secondly, we confirm that effective local modeling is essential for the success of Softmax attention, in which linear attention falls short. The aforementioned two fundamental differences significantly contribute to the disparities between these two attention paradigms, which is demonstrated by our substantial empirical validation in the paper. In addition, more experiment results indicate that linear attention, as long as endowed with these two properties, can outperform Softmax attention across various tasks while maintaining lower computation complexity. Code is available at https://github.com/LeapLabTHU/InLine.

Although transformer-based methods have achieved great success in multi-scale temporal pattern interaction modeling, two key challenges limit their further development: (1) Individual time points contain less semantic information, and leveraging attention to model pair-wise interactions may cause the information utilization bottleneck. (2) Multiple inherent temporal variations (e.g., rising, falling, and fluctuating) entangled in temporal patterns. To this end, we propose Adaptive Multi-Scale Hypergraph Transformer (Ada-MSHyper) for time series forecasting. Specifically, an adaptive hypergraph learning module is designed to provide foundations for modeling group-wise interactions, then a multi-scale interaction module is introduced to promote more comprehensive pattern interactions at different scales. In addition, a node and hyperedge constraint mechanism is introduced to cluster nodes with similar semantic information and differentiate the temporal variations within each scales. Extensive experiments on 11 real-world datasets demonstrate that Ada-MSHyper achieves state-of-the-art performance, reducing prediction errors by an average of 4.56%, 10.38%, and 4.97% in MSE for long-range, short-range, and ultra-long-range time series forecasting, respectively. Code is available at https://github.com/shangzongjiang/Ada-MSHyper.

Out-of-distribution (OOD) generalization issue is a well-known challenge within deep learning tasks. In dynamic graphs, the change of temporal environments is regarded as the main cause of data distribution shift. While numerous OOD studies focusing on environment factors have achieved remarkable performance, they still fail to systematically solve the two issue of environment inference and utilization. In this work, we propose a novel dynamic graph learning model named EpoD based on prompt learning and structural causal model to comprehensively enhance both environment inference and utilization. Inspired by the superior performance of prompt learning in understanding underlying semantic and causal associations, we first design a self-prompted learning mechanism to infer unseen environment factors. We then rethink the role of environment variable within spatio-temporal causal structure model, and introduce a novel causal pathway where dynamic subgraphs serve as mediating variables. The extracted dynamic subgraph can effectively capture the data distribution shift by incorporating the inferred environment variables into the node-wise dependencies. Theoretical discussions and intuitive analysis support the generalizability and interpretability of EpoD. Extensive experiments on seven real-world datasets across domains showcase the superiority of EpoD against baselines, and toy example experiments further verify the powerful interpretability and rationality of our EpoD.

Multimodal learning falls into the trap of the optimization dilemma due to the modality imbalance phenomenon, leading to unsatisfactory performance in real applications. A core reason for modality imbalance is that the models of each modality converge at different rates. Many attempts naturally focus on adjusting learning procedures adaptively. Essentially, the reason why models converge at different rates is because the difficulty of fitting category labels is inconsistent for each modality during learning. From the perspective of fitting labels, we find that appropriate positive intervention label fitting can correct this difference in learning ability. By exploiting the ability of contrastive learning to intervene in the learning of category label fitting, we propose a novel multimodal learning approach that dynamically integrates unsupervised contrastive learning and supervised multimodal learning to address the modality imbalance problem. We find that a simple yet heuristic integration strategy can significantly alleviate the modality imbalance phenomenon. Moreover, we design a learning-based integration strategy to integrate two losses dynamically, further improving the performance. Experiments on widely used datasets demonstrate the superiority of our method compared with state-of-the-art (SOTA) multimodal learning approaches. The code is available at https://github.com/njustkmg/NeurIPS24-LFM.

The attention mechanism within the transformer architecture enables the model to weigh and combine tokens based on their relevance to the query. While self-attention has enjoyed major success, it notably treats all queries $q$ in the same way by applying the mapping $V^\top\text{softmax}(Kq)$, where $V,K$ are the value and key embeddings respectively. In this work, we argue that this uniform treatment hinders the ability to control contextual sparsity and relevance. As a solution, we introduce the Selective Self-Attention (SSA) layer that augments the softmax nonlinearity with a principled temperature scaling strategy. By controlling temperature, SSA adapts the contextual sparsity of the attention map to the query embedding and its position in the context window. Through theory and experiments, we demonstrate that this alleviates attention dilution, aids the optimization process, and enhances the model's ability to control softmax spikiness of individual queries. We also incorporate temperature scaling for value embeddings and show that it boosts the model's ability to suppress irrelevant/noisy tokens. Notably, SSA is a lightweight method which introduces less than 0.5\% new parameters through a weight-sharing strategy and can be fine-tuned on existing LLMs. Extensive empirical evaluations demonstrate that SSA-equipped models achieve a noticeable and consistent accuracy improvement on language modeling benchmarks.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable comprehension and reasoning capabilities with complex language and visual data. These advances have spurred the vision of establishing a generalist robotic MLLM proficient in understanding complex human instructions and accomplishing various embodied tasks, whose feasibility has been recently verified~\cite{rt-2,rt-x}. However, developing MLLMs for real-world robots is challenging due to the typically limited computation and memory capacities available on robotic platforms. In contrast, the inference of MLLMs usually incorporates storing billions of parameters and performing tremendous computation, imposing significant hardware demands. In our paper, we seek to address this challenge by leveraging an intriguing observation: relatively easier situations make up the bulk of the procedure of controlling robots to fulfill diverse tasks, and they generally require far smaller models to obtain the correct robotic actions. Motivated by this observation, we propose a \emph{Dynamic Early-Exit for Robotic MLLM} (DeeR) framework that automatically adjusts the size of the activated MLLM based on each situation at hand. The approach leverages a multi-exit architecture in MLLMs, which allows the model to cease processing once a proper size of the model has been activated for a specific situation, thus avoiding further redundant computation. Additionally, we develop novel algorithms that establish early-termination criteria for DeeR, conditioned on predefined demands such as average computational cost (\emph{i.e.}, power consumption), as well as peak computational consumption (\emph{i.e.}, latency) and GPU memory usage. These enhancements ensure that DeeR operates efficiently under varying resource constraints while maintaining competitive performance. Moreover, we design a tailored training method for integrating temporal information on top of such multi-exit architectures to predict actions reasonably. On the CALVIN robot manipulation benchmark, DeeR demonstrates significant reductions in computational costs by 5.2-6.5x and GPU memory by 2x without compromising performance. Code and checkpoints are available at https://github.com/yueyang130/DeeR-VLA.

This paper reveals the phenomenon of parameter heterogeneity in large language models (LLMs). We find that a small subset of ``cherry'' parameters exhibit a disproportionately large influence on model performance, while the vast majority of parameters have minimal impact. This heterogeneity is found to be prevalent across different model families, scales, and types. Motivated by this observation, we propose CherryQ, a novel quantization method that unifies the optimization of mixed-precision parameters. CherryQ identifies and preserves the critical cherry parameters in high precision while aggressively quantizing the remaining parameters to low precision. Extensive experiments demonstrate the effectiveness of CherryQ. CherryQ outperforms existing quantization approaches in terms of perplexity and downstream task performance. Notably, our 3-bit quantized Vicuna-1.5 exhibits competitive performance compared to their 16-bit counterparts. These findings highlight the potential of CherryQ for enabling efficient deployment of LLMs by taking advantage of parameter heterogeneity.

The problem of heterogeneous clients in federated learning has recently drawn a lot of attention. Spectral model sharding, i.e., partitioning the model parameters into low-rank matrices based on the singular value decomposition, has been one of the proposed solutions for more efficient on-device training in such settings. In this work we present two sampling strategies for such sharding, obtained as solutions to specific optimization problems. The first produces unbiased estimators of the original weights, while the second aims to minimize the squared approximation error. We discuss how both of these estimators can be incorporated in the federated learning loop and practical considerations that arise during local training. Empirically, we demonstrate that both of these methods can lead to improved performance in various commonly used datasets.

We introduce a novel positional encoding strategy for Transformer-style models, addressing the shortcomings of existing, often ad hoc, approaches. Our framework implements a flexible mapping from the algebraic specification of a domain to a positional encoding scheme where positions are interpreted as orthogonal operators. This design preserves the structural properties of the source domain, thereby ensuring that the end-model upholds them. The framework can accommodate various structures, including sequences, grids and trees, but also their compositions. We conduct a series of experiments demonstrating the practical applicability of our method. Our results suggest performance on par with or surpassing the current state of the art, without hyper-parameter optimizations or ``task search'' of any kind. Code is available through https://aalto-quml.github.io/ape/.

Increasing the throughput of the Transformer architecture, a foundational component used in numerous state-of-the-art models for vision and language tasks (e.g., GPT, LLaVa), is an important problem in machine learning. One recent and effective strategy is to merge token representations within Transformer models, aiming to reduce computational and memory requirements while maintaining accuracy. Prior work has proposed algorithms based on Bipartite Soft Matching (BSM), which divides tokens into distinct sets and merges the top $k$ similar tokens. However, these methods have significant drawbacks, such as sensitivity to token-splitting strategies and damage to informative tokens in later layers. This paper presents a novel paradigm called PiToMe, which prioritizes the preservation of informative tokens using an additional metric termed the \textit{energy score}. This score identifies large clusters of similar tokens as high-energy, indicating potential candidates for merging, while smaller (unique and isolated) clusters are considered as low-energy and preserved. Experimental findings demonstrate that PiToMe saved from 40-60\% FLOPs of the base models while exhibiting superior off-the-shelf performance on image classification (0.5\% average performance drop of ViT-MAEH compared to 2.6\% as baselines), image-text retrieval (0.3\% average performance drop of Clip on Flick30k compared to 4.5\% as others), and analogously in visual questions answering with LLaVa-7B. Furthermore, PiToMe is theoretically shown to preserve intrinsic spectral properties to the original token space under mild conditions.