NeurIPS 2024 论文列表

High-dimensional data commonly lies on low-dimensional submanifolds, and estimating the local intrinsic dimension (LID) of a datum -- i.e. the dimension of the submanifold it belongs to -- is a longstanding problem. LID can be understood as the number of local factors of variation: the more factors of variation a datum has, the more complex it tends to be. Estimating this quantity has proven useful in contexts ranging from generalization in neural networks to detection of out-of-distribution data, adversarial examples, and AI-generated text. The recent successes of deep generative models present an opportunity to leverage them for LID estimation, but current methods based on generative models produce inaccurate estimates, require more than a single pre-trained model, are computationally intensive, or do not exploit the best available deep generative models: diffusion models (DMs). In this work, we show that the Fokker-Planck equation associated with a DM can provide an LID estimator which addresses the aforementioned deficiencies. Our estimator, called FLIPD, is easy to implement and compatible with all popular DMs. Applying FLIPD to synthetic LID estimation benchmarks, we find that DMs implemented as fully-connected networks are highly effective LID estimators that outperform existing baselines. We also apply FLIPD to natural images where the true LID is unknown. Despite being sensitive to the choice of network architecture, FLIPD estimates remain a useful measure of relative complexity; compared to competing estimators, FLIPD exhibits a consistently higher correlation with image PNG compression rate and better aligns with qualitative assessments of complexity. Notably, FLIPD is orders of magnitude faster than other LID estimators, and the first to be tractable at the scale of Stable Diffusion.

Large-scale foundation models have demonstrated exceptional performance in language and vision tasks. However, the numerous dense matrix-vector operations involved in these large networks pose significant computational challenges during inference. To address these challenges, we introduce the Block-Level Adaptive STructured (BLAST) matrix, designed to learn and leverage efficient structures prevalent in the weight matrices of linear layers within deep learning models. Compared to existing structured matrices, the BLAST matrix offers substantial flexibility, as it can represent various types of structures that are either learned from data or computed from pre-existing weight matrices. We demonstrate the efficiency of using the BLAST matrix for compressing both language and vision tasks, showing that (i) for medium-sized models such as ViT and GPT-2, training with BLAST weights boosts performance while reducing complexity by 70\% and 40\%, respectively; and (ii) for large foundation models such as Llama-7B and DiT-XL, the BLAST matrix achieves a 2x compression while exhibiting the lowest performance degradation among all tested structured matrices. Our code is available at https://github.com/changwoolee/BLAST.

Remote sensing (RS) imagery, which requires specialized satellites to collect and is difficult to annotate, suffers from data scarcity and class imbalance in certain spectrums. Due to their data scarcity, training large-scale RS models from scratch is unrealistic, and the alternative is to transfer pre-trained models by fine-tuning or a more data-efficient method LoRA. Due to class imbalance, transferred models exhibit strong bias, where features of the major class dominate over those of the minor class. In this paper, we propose debLoRA, a generic training approach that works with any LoRA variants to yield debiased features. It is an unsupervised learning approach that can diversify minor class features based on the shared attributes with major classes, where the attributes are obtained by a simple step of clustering. To evaluate it, we conduct extensive experiments in two transfer learning scenarios in the RS domain: from natural to optical RS images, and from optical RS to multi-spectrum RS images. We perform object classification and oriented object detection tasks on the optical RS dataset DOTA and the SAR dataset FUSRS. Results show that our debLoRA consistently surpasses prior arts across these RS adaptation settings, yielding up to 3.3 and 4.7 percentage points gains on the tail classes for natural $\to$ optical RS and optical RS $\to$ multi-spectrum RS adaptations, respectively, while preserving the performance on head classes, substantiating its efficacy and adaptability

Bounds on the smallest eigenvalue of the neural tangent kernel (NTK) are a key ingredient in the analysis of neural network optimization and memorization. However, existing results require distributional assumptions on the data and are limited to a high-dimensional setting, where the input dimension $d_0$ scales at least logarithmically in the number of samples $n$. In this work we remove both of these requirements and instead provide bounds in terms of a measure of distance between data points: notably these bounds hold with high probability even when $d_0$ is held constant versus $n$. We prove our results through a novel application of the hemisphere transform.

Large transformer-based foundation models have been commonly used as pre-trained models that can be adapted to different challenging datasets and settings with state-of-the-art generalization performance. Parameter efficient fine-tuning ($\texttt{PEFT}$) provides promising generalization performance in adaptation while incurring minimum computational overhead. However, adaptation of these foundation models through $\texttt{PEFT}$ leads to accurate but severely underconfident models, especially in few-shot learning settings. Moreover, the adapted models lack accurate fine-grained uncertainty quantification capabilities limiting their broader applicability in critical domains. To fill out this critical gap, we develop a novel lightweight {Bayesian Parameter Efficient Fine-Tuning} (referred to as $\texttt{Bayesian-PEFT}$) framework for large transformer-based foundation models. The framework integrates state-of-the-art $\texttt{PEFT}$ techniques with two Bayesian components to address the under-confidence issue while ensuring reliable prediction under challenging few-shot settings. The first component performs base rate adjustment to strengthen the prior belief corresponding to the knowledge gained through pre-training, making the model more confident in its predictions; the second component builds an evidential ensemble that leverages belief regularization to ensure diversity among different ensemble components. Our thorough theoretical analysis justifies that the Bayesian components can ensure reliable and accurate few-shot adaptations with well-calibrated uncertainty quantification. Extensive experiments across diverse datasets, few-shot learning scenarios, and multiple $\texttt{PEFT}$ techniques demonstrate the outstanding prediction and calibration performance by $\texttt{Bayesian-PEFT}$.

The prohibitive sizes of Large Language Models (LLMs) today make it difficult to deploy them on memory-constrained edge devices. This work introduces $\rm CALDERA$ -- a new post-training LLM compression algorithm that harnesses the inherent low-rank structure of a weight matrix $\mathbf{W}$ by approximating it via a low-rank, low-precision decomposition as $\mathbf{W} \approx \mathbf{Q} + \mathbf{L}\mathbf{R}$. Here, $\mathbf{L}$ and $\mathbf{R}$ are low rank factors, and the entries of $\mathbf{Q}$, $\mathbf{L}$ and $\mathbf{R}$ are quantized. The model is compressed by substituting each layer with its $\mathbf{Q} + \mathbf{L}\mathbf{R}$ decomposition, and the zero-shot performance of the compressed model is evaluated. Additionally, $\mathbf{L}$ and $\mathbf{R}$ are readily amenable to low-rank adaptation, consequently enhancing the zero-shot performance. $\rm CALDERA$ obtains this decomposition by formulating it as an optimization problem $\min_{\mathbf{Q},\mathbf{L},\mathbf{R}}\lVert(\mathbf{Q} + \mathbf{L}\mathbf{R} - \mathbf{W})\mathbf{X}^\top\rVert_{\rm F}^2$, where $\mathbf{X}$ is the calibration data, and $\mathbf{Q}, \mathbf{L}, \mathbf{R}$ are constrained to be representable using low-precision formats. Theoretical upper bounds on the approximation error of $\rm CALDERA$ are established using a rank-constrained regression framework, and the tradeoff between compression ratio and model performance is studied by analyzing the impact of target rank and quantization bit budget. Results illustrate that compressing LlaMa-$2$ $7$B/$13$B/$70$B and LlaMa-$3$ $8$B models obtained using $\rm CALDERA$ outperforms existing post-training LLM compression techniques in the regime of less than $2.5$ bits per parameter.

Recent research has shown that Transformers with linear attention are capable of in-context learning (ICL) by implementing a linear estimator through gradient descent steps. However, the existing results on the optimization landscape apply under stylized settings where task and feature vectors are assumed to be IID and the attention weights are fully parameterized. In this work, we develop a stronger characterization of the optimization and generalization landscape of ICL through contributions on architectures, low-rank parameterization, and correlated designs: (1) We study the landscape of 1-layer linear attention and 1-layer H3, a state-space model. Under a suitable correlated design assumption, we prove that both implement 1-step preconditioned gradient descent. We show that thanks to its native convolution filters, H3 also has the advantage of implementing sample weighting and outperforming linear attention in suitable settings. (2) By studying correlated designs, we provide new risk bounds for retrieval augmented generation (RAG) and task-feature alignment which reveal how ICL sample complexity benefits from distributional alignment. (3) We derive the optimal risk for low-rank parameterized attention weights in terms of covariance spectrum. Through this, we also shed light on how LoRA can adapt to a new distribution by capturing the shift between task covariances. Experimental results corroborate our theoretical findings. Overall, this work explores the optimization and risk landscape of ICL in practically meaningful settings and contributes to a more thorough understanding of its mechanics.

Current PEFT methods for LLMs can achieve high quality, efficient training, or scalable serving, but not all three simultaneously. To address this limitation, we investigate sparse fine-tuning and observe a remarkable improvement in generalization ability. Utilizing this key insight, we propose a family of Structured Sparse Fine-Tuning (S${^2}$FT) methods for LLMs, which concurrently achieve state-of-the-art fine-tuning performance, training efficiency, and inference scalability. S${^2}$FT accomplishes this by "selecting sparsely and computing densely". Based on the coupled structures in LLMs, \model selects a few attention heads and channels in the MHA and FFN modules for each Transformer block, respectively. Next, it co-permutes the weight matrices on both sides of all coupled structures to connect the selected subsets in each layer into a dense submatrix. Finally, S${^2}$FT performs in-place gradient updates on all selected submatrices. Through theoretical analyses and empirical results, our method prevents forgetting while simplifying optimization, delivers SOTA performance on both commonsense and arithmetic reasoning with 4.6% and 1.3% average improvements compared to LoRA, and surpasses full FT by 11.5% when generalizing to various domains after instruction tuning. Using our partial back-propagation algorithm, S${^2}$FT saves training memory up to 3$\times$ and improves latency by 1.5-2.7$\times$ compared to full FT, while achieving an average 10\% improvement over LoRA on both metrics. We further demonstrate that the weight updates in S${^2}$FT can be decoupled into adapters, enabling effective fusion, fast switch, and efficient parallelism when serving multiple fine-tuned models.

Recent advances in low light image enhancement have been dominated by Retinex-based learning framework, leveraging convolutional neural networks (CNNs) and Transformers. However, the vanilla Retinex theory primarily addresses global illumination degradation and neglects local issues such as noise and blur in dark conditions. Moreover, CNNs and Transformers struggle to capture global degradation due to their limited receptive fields. While state space models (SSMs) have shown promise in the long-sequence modeling, they face challenges in combining local invariants and global context in visual data. In this paper, we introduce MambaLLIE, an implicit Retinex-aware low light enhancer featuring a global-then-local state space design. We first propose a Local-Enhanced State Space Module (LESSM) that incorporates an augmented local bias within a 2D selective scan mechanism, enhancing the original SSMs by preserving local 2D dependency. Additionally, an Implicit Retinex-aware Selective Kernel module (IRSK) dynamically selects features using spatially-varying operations, adapting to varying inputs through an adaptive kernel selection process. Our Global-then-Local State Space Block (GLSSB) integrates LESSM and IRSK with layer normalization (LN) as its core. This design enables MambaLLIE to achieve comprehensive global long-range modeling and flexible local feature aggregation. Extensive experiments demonstrate that MambaLLIE significantly outperforms state-of-the-art CNN and Transformer-based methods. Our code is available at https://github.com/wengjiangwei/MambaLLIE.

While fine-tuning pretrained models has become common practice, these models often underperform outside their specific domains. Recently developed model merging techniques enable the direct integration of multiple models, each fine-tuned for distinct tasks, into a single model. This strategy promotes multitasking capabilities without requiring retraining on the original datasets. However, existing methods fall short in addressing potential conflicts and complex correlations between tasks, especially in parameter-level adjustments, posing a challenge in effectively balancing parameter competition across various tasks. This paper introduces an innovative technique named **PCB-Merging** (Parameter Competition Balancing), a *lightweight* and *training-free* technique that adjusts the coefficients of each parameter for effective model merging. PCB-Merging employs intra-balancing to gauge parameter significance within individual tasks and inter-balancing to assess parameter similarities across different tasks. Parameters with low importance scores are dropped, and the remaining ones are rescaled to form the final merged model. We assessed our approach in diverse merging scenarios, including cross-task, cross-domain, and cross-training configurations, as well as out-of-domain generalization. The experimental results reveal that our approach achieves substantial performance enhancements across multiple modalities, domains, model sizes, number of tasks, fine-tuning forms, and large language models, outperforming existing model merging methods.

As the adoption of large language models increases and the need for per-user or per-task model customization grows, the parameter-efficient fine-tuning (PEFT) methods, such as low-rank adaptation (LoRA) and its variants, incur substantial storage and transmission costs. To further reduce stored parameters, we introduce a "divide-and-share" paradigm that breaks the barriers of low-rank decomposition across matrix dimensions, modules, and layers by sharing parameters globally via a vector bank. As an instantiation of the paradigm to LoRA, our proposed VB-LoRA composites all the low-rank matrices of LoRA from a shared vector bank with a differentiable top-$k$ admixture module. VB-LoRA achieves extreme parameter efficiency while maintaining comparable or better performance compared to state-of-the-art PEFT methods. Extensive experiments demonstrate the effectiveness of VB-LoRA on natural language understanding, natural language generation, instruction tuning, and mathematical reasoning tasks. When fine-tuning the Llama2-13B model, VB-LoRA only uses 0.4% of LoRA's stored parameters, yet achieves superior results. Our source code is available at https://github.com/leo-yangli/VB-LoRA. This method has been merged into the Hugging Face PEFT package.

Text embeddings are essential for tasks such as document retrieval, clustering, and semantic similarity assessment. In this paper, we study how to contrastively train text embedding models in a compute-optimal fashion, given a suite of pretrained decoder-only language models. Our innovation is an algorithm that produces optimal configurations of model sizes, data quantities, and fine-tuning methods for text-embedding models at different computational budget levels. The resulting recipe, which we obtain through extensive experiments, can be used by practitioners to make informed design choices for their embedding models. Specifically, our findings suggest that full fine-tuning and Low-Rank Adaptation fine-tuning produce optimal models at lower and higher computational budgets respectively.

We introduce a new linear-algebraic tool based on group representation theory, and use it to address three key problems in machine learning. 1. Past researchers have proposed fast attention algorithms for LLMs by approximating or replace softmax attention with other functions, such as low-degree polynomials. The key property of these functions is that, when applied entry-wise to the matrix $QK^{\top}$, the result is a low rank matrix when $Q$ and $K$ are $n \times d$ matrices and $n \gg d$. This suggests a natural question: what are all functions $f$ with this property? If other $f$ exist and are quickly computable, they can be used in place of softmax for fast subquadratic attention algorithms. It was previously known that low-degree polynomials have this property. We prove that low-degree polynomials are the only piecewise continuous functions with this property. This suggests that the low-rank fast attention only works for functions approximable by polynomials. Our work gives a converse to the polynomial method in algorithm design. 2. We prove the first full classification of all positive definite kernels that are functions of Manhattan or $\ell_1$ distance. Our work generalizes an existing theorem at the heart of all kernel methods in machine learning: the classification of all positive definite kernels that are functions of Euclidean distance. 3. The key problem in metric transforms, a mathematical theory used in geometry and machine learning, asks what functions transform pairwise distances in semi-metric space $M$ to semi-metric space $N$ for specified $M$ and $N$. We provide the first full classification of functions that transform Manhattan distances to Manhattan distances. Our work generalizes the foundational work of Schoenberg, which fully classifies functions that transform Euclidean to Euclidean distances. We additionally prove results about stable-rank preserving functions that are potentially useful in algorithmic design, and more. Our core new tool is called the representation theory of the hyperrectangle.

Training a policy in a source domain for deployment in the target domain under a dynamics shift can be challenging, often resulting in performance degradation. Previous work tackles this challenge by training on the source domain with modified rewards derived by matching distributions between the source and the target optimal trajectories. However, pure modified rewards only ensure the behavior of the learned policy in the source domain resembles trajectories produced by the target optimal policies, which does not guarantee optimal performance when the learned policy is actually deployed to the target domain. In this work, we propose to utilize imitation learning to transfer the policy learned from the reward modification to the target domain so that the new policy can generate the same trajectories in the target domain. Our approach, Domain Adaptation and Reward Augmented Imitation Learning (DARAIL), utilizes the reward modification for domain adaptation and follows the general framework of generative adversarial imitation learning from observation (GAIfO) by applying a reward augmented estimator for the policy optimization step. Theoretically, we present an error bound for our method under a mild assumption regarding the dynamics shift to justify the motivation of our method. Empirically, our method outperforms the pure modified reward method without imitation learning and also outperforms other baselines in benchmark off-dynamics environments.

Spiking neural networks (SNNs), inspired by biological processes, use spike signals for inter-layer communication, presenting an energy-efficient alternative to traditional neural networks. To realize the theoretical advantages of SNNs in energy efficiency, it is essential to deploy them onto neuromorphic chips. On clock-driven synchronous chips, employing shorter time steps can enhance energy efficiency but reduce SNN performance. Compared to the clock-driven synchronous chip, the event-driven asynchronous chip achieves much lower energy consumption but only supports some specific network operations. Recently, a series of SNN projects have achieved tremendous success, significantly improving the SNN's performance. However, event-driven asynchronous chips do not support some of the proposed structures, making it impossible to integrate these SNNs into asynchronous hardware. In response to these problems, we propose the Spiking Token Mixer (STMixer) architecture, which consists exclusively of operations supported by asynchronous scenarios, including convolutional, fully connected layers and residual paths. Our series of experiments also demonstrates that STMixer achieves performance on par with spiking transformers in synchronous scenarios with very low timesteps. This indicates its ability to achieve the same level of performance with lower power consumption in synchronous scenarios. The codes are available at \url{https://github.com/brain-intelligence-lab/STMixer_demo}.

In recent years, indoor air pollution has posed a significant threat to our society, claiming over 3.2 million lives annually. Developing nations, such as India, are most affected since lack of knowledge, inadequate regulation, and outdoor air pollution lead to severe daily exposure to pollutants. However, only a limited number of studies have attempted to understand how indoor air pollution affects developing countries like India. To address this gap, we present spatiotemporal measurements of air quality from 30 indoor sites over six months during summer and winter seasons. The sites are geographically located across four regions of type: rural, suburban, and urban, covering the typical low to middle-income population in India. The dataset contains various types of indoor environments (e.g., studio apartments, classrooms, research laboratories, food canteens, and residential households), and can provide the basis for data-driven learning model research aimed at coping with unique pollution patterns in developing countries. This unique dataset demands advanced data cleaning and imputation techniques for handling missing data due to power failure or network outages during data collection. Furthermore, through a simple speech-to-text application, we provide real-time indoor activity labels annotated by occupants. Therefore, environmentalists and ML enthusiasts can utilize this dataset to understand the complex patterns of the pollutants under different indoor activities, identify recurring sources of pollution, forecast exposure, improve floor plans and room structures of modern indoor designs, develop pollution-aware recommender systems, etc.

Training large AI models such as LLMs and DLRMs costs massive GPUs and computing time. The high training cost has become only affordable to big tech companies, meanwhile also causing increasing concerns about the environmental impact. This paper presents CoMERA, a **Co**mputing- and **M**emory-**E**fficient training method via **R**ank-**A**daptive tensor optimization. CoMERA achieves end-to-end rank-adaptive tensor-compressed training via a multi-objective optimization formulation, and improves the training to provide both a high compression ratio and excellent accuracy in the training process. Our optimized numerical computation (e.g., optimized tensorized embedding and tensor-vector contractions) and GPU implementation eliminate part of the run-time overhead in the tensorized training on GPU. This leads to, for the first time, $2-3\times$ speedup per training epoch compared with standard training. CoMERA also outperforms the recent GaLore in terms of both memory and computing efficiency. Specifically, CoMERA is $2\times$ faster per training epoch and $9\times$ more memory-efficient than GaLore on a tested six-encoder transformer with single-batch training. Our method also shows $\sim 2\times$ speedup than standard pre-training on a BERT-like code-generation LLM while achieving $4.23\times$ compression ratio in pre-training. With further HPC optimization, CoMERA may reduce the pre-training cost of many other LLMs. An implementation of CoMERA is available at <https://github.com/ziyangjoy/CoMERA>.

Optimal transport (OT) is a general framework for finding a minimum-cost transport plan, or coupling, between probability distributions, and has many applications in machine learning. A key challenge in applying OT to massive datasets is the quadratic scaling of the coupling matrix with the size of the dataset. [Forrow et al. 2019] introduced a factored coupling for the k-Wasserstein barycenter problem, which [Scetbon et al. 2021] adapted to solve the primal low-rank OT problem. We derive an alternative parameterization of the low-rank problem based on the _latent coupling_ (LC) factorization previously introduced by [Lin et al. 2021] generalizing [Forrow et al. 2019]. The LC factorization has multiple advantages for low-rank OT including decoupling the problem into three OT problems and greater flexibility and interpretability. We leverage these advantages to derive a new algorithm _Factor Relaxation with Latent Coupling_ (FRLC), which uses _coordinate_ mirror descent to compute the LC factorization. FRLC handles multiple OT objectives (Wasserstein, Gromov-Wasserstein, Fused Gromov-Wasserstein), and marginal constraints (balanced, unbalanced, and semi-relaxed) with linear space complexity. We provide theoretical results on FRLC, and demonstrate superior performance on diverse applications -- including graph clustering and spatial transcriptomics -- while demonstrating its interpretability.

Convolutional residual neural networks (ConvResNets), though overparametersized, can achieve remarkable prediction performance in practice, which cannot be well explained by conventional wisdom. To bridge this gap, we study the performance of ConvResNeXts trained with weight decay, which cover ConvResNets as a special case, from the perspective of nonparametric classification. Our analysis allows for infinitely many building blocks in ConvResNeXts, and shows that weight decay implicitly enforces sparsity on these blocks. Specifically, we consider a smooth target function supported on a low-dimensional manifold, then prove that ConvResNeXts can adapt to the function smoothness and low-dimensional structures and efficiently learn the function without suffering from the curse of dimensionality. Our findings partially justify the advantage of overparameterized ConvResNeXts over conventional machine learning models.