NeurIPS 2024 Papers

Sharpness-aware minimization (SAM) improves generalization of various deep learning tasks. Motivated by popular architectures such as LoRA, we explore the implicit regularization of SAM for scale-invariant problems involving two groups of variables. Instead of focusing on commonly used sharpness, this work introduces a concept termed *balancedness*, defined as the difference between the squared norm of two variables. This allows us to depict richer global behaviors of SAM. In particular, our theoretical and empirical findings reveal that i) SAM promotes balancedness; and ii) the regularization on balancedness is *data-responsive* -- outliers have stronger impact. The latter coincides with empirical observations that SAM outperforms SGD in the presence of outliers. Leveraging the implicit regularization, we develop a resource-efficient SAM variant, balancedness-aware regularization (BAR), tailored for scale-invariant problems such as finetuning language models with LoRA. BAR saves 95% computational overhead of SAM, with enhanced test performance across various tasks on RoBERTa, GPT2, and OPT-1.3B.

This paper investigates simple bilevel optimization problems where we minimize a convex upper-level objective over the optimal solution set of a convex lower-level objective. Existing methods for such problems either only guarantee asymptotic convergence, have slow sublinear rates, or require strong assumptions. To address these challenges, we propose a penalization framework that delineates the relationship between approximate solutions of the original problem and its reformulated counterparts. This framework accommodates varying assumptions regarding smoothness and convexity, enabling the application of specific methods with different complexity results. Specifically, when both upper- and lower-level objectives are composite convex functions, under an $\alpha$-Hölderian error bound condition and certain mild assumptions, our algorithm attains an $(\epsilon,\epsilon^{\beta})$-optimal solution of the original problem for any $\beta> 0$ within $\mathcal{O}\left(\sqrt{{1}/{\epsilon^{\max\\{\alpha,\beta\\}}}}\right)$ iterations. The result can be improved further if the smooth part of the upper-level objective is strongly convex. We also establish complexity results when the upper- and lower-level objectives are general nonsmooth functions. Numerical experiments demonstrate the effectiveness of our algorithms.

The effect of regularizers such as weight decay when training deep neural networks is not well understood. We study the influence of weight decay as well as $L2$-regularization when training neural network models in which parameter matrices interact multiplicatively. This combination is of particular interest as this parametrization is common in attention layers, the workhorse of transformers. Here, key-query, as well as value-projection parameter matrices, are multiplied directly with each other: $W_K^TW_Q$ and $PW_V$. We extend previous results and show on one hand that any local minimum of a $L2$-regularized loss of the form $L(AB^\top) + \lambda (\|A\|^2 + \|B\|^2)$ coincides with a minimum of the nuclear norm-regularized loss $L(AB^\top) + \lambda\|AB^\top\|_*$, and on the other hand that the 2 losses become identical exponentially quickly during training. We thus complement existing works linking $L2$-regularization with low-rank regularization, and in particular, explain why such regularization on the matrix product affects early stages of training. Based on these theoretical insights, we verify empirically that the key-query and value-projection matrix products $W_K^TW_Q, PW_V$ within attention layers, when optimized with weight decay, as usually done in vision tasks and language modelling, indeed induce a significant reduction in the rank of $W_K^TW_Q$ and $PW_V$, even in fully online training. We find that, in accordance with existing work, inducing low rank in attention matrix products can damage language model performance, and observe advantages when decoupling weight decay in attention layers from the rest of the parameters.

Second-order optimization algorithms, such as the Newton method and the natural gradient descent (NGD) method exhibit excellent convergence properties for training deep neural networks, but the high computational cost limits its practical application. In this paper, we focus on the NGD method and propose a novel layer-wise natural gradient descent (LNGD) method to further reduce computational costs and accelerate the training process. Specifically, based on the block diagonal approximation of the Fisher information matrix, we first propose the layer-wise sample method to compute each block matrix without performing a complete back-propagation. Then, each block matrix is approximated as a Kronecker product of two smaller matrices, one of which is a diagonal matrix, while keeping the traces equal before and after approximation. By these two steps, we provide a new approximation for the Fisher information matrix, which can effectively reduce the computational cost while preserving the main information of each block matrix. Moreover, we propose a new adaptive layer-wise learning rate to further accelerate training. Based on these new approaches, we propose the LNGD optimizer. The global convergence analysis of LNGD is established under some assumptions. Experiments on image classification and machine translation tasks show that our method is quite competitive compared to the state-of-the-art methods.

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.

Dataset condensation (DC) is an emerging technique capable of creating compact synthetic datasets from large originals while maintaining considerable performance. It is crucial for accelerating network training and reducing data storage requirements. However, current research on DC mainly focuses on image classification, with less exploration of object detection. This is primarily due to two challenges: (i) the multitasking nature of object detection complicates the condensation process, and (ii) Object detection datasets are characterized by large-scale and high-resolution data, which are difficult for existing DC methods to handle. As a remedy, we propose DCOD, the first dataset condensation framework for object detection. It operates in two stages: Fetch and Forge, initially storing key localization and classification information into model parameters, and then reconstructing synthetic images via model inversion. For the complex of multiple objects in an image, we propose Foreground Background Decoupling to centrally update the foreground of multiple instances and Incremental PatchExpand to further enhance the diversity of foregrounds. Extensive experiments on various detection datasets demonstrate the superiority of DCOD. Even at an extremely low compression rate of 1\%, we achieve 46.4\% and 24.7\% $\text{AP}_{50}$ on the VOC and COCO, respectively, significantly reducing detector training duration.

This paper aims at developing novel shuffling gradient-based methods for tackling two classes of minimax problems: nonconvex-linear and nonconvex-strongly concave settings. The first algorithm addresses the nonconvex-linear minimax model and achieves the state-of-the-art oracle complexity typically observed in nonconvex optimization. It also employs a new shuffling estimator for the ``hyper-gradient'', departing from standard shuffling techniques in optimization. The second method consists of two variants: semi-shuffling and full-shuffling schemes. These variants tackle the nonconvex-strongly concave minimax setting. We establish their oracle complexity bounds under standard assumptions, which, to our best knowledge, are the best-known for this specific setting. Numerical examples demonstrate the performance of our algorithms and compare them with two other methods. Our results show that the new methods achieve comparable performance with SGD, supporting the potential of incorporating shuffling strategies into minimax algorithms.

The success of pretrain-finetune paradigm brings about the release of numerous model weights. In this case, merging models finetuned on different tasks to enable a single model with multi-task capabilities is gaining increasing attention for its practicability. Existing model merging methods usually suffer from (1) significant performance degradation or (2) requiring tuning by additional data or training. In this paper, we rethink and analyze the existing model merging paradigm. We discover that using a single model's weights can hardly simulate all the models' performance. To tackle this issue, we propose Elect, Mask & Rescale-Merging (EMR-Merging). We first (a) elect a unified model from all the model weights and then (b) generate extremely lightweight task-specific modulators, including masks and rescalers, to align the direction and magnitude between the unified model and each specific model, respectively. EMR-Merging is tuning-free, thus requiring no data availability or any additional training while showing impressive performance. We find that EMR-Merging shows outstanding performance compared to existing merging methods under different classical and newly-established settings, including merging different numbers of vision models (up to 30), NLP models, PEFT models, and multi-modal models.

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.

Fine-tuning pre-trained models for downstream tasks is a widely adopted technique known for its adaptability and reliability across various domains. Despite its conceptual simplicity, fine-tuning entails several troublesome engineering choices, such as selecting hyperparameters and determining checkpoints from an optimization trajectory. To tackle the difficulty of choosing the best model, one effective solution is model fusion, which combines multiple models in a parameter space. However, we observe a large discrepancy between loss and metric landscapes during the fine-tuning of pre-trained language models. Building on this observation, we introduce a novel model fusion technique that optimizes both the desired metric and loss through multi-objective Bayesian optimization. In addition, to effectively select hyperparameters, we establish a two-stage procedure by integrating Bayesian optimization processes into our framework. Experiments across various downstream tasks show considerable performance improvements using our Bayesian optimization-guided method.

Contextual linear optimization (CLO) uses predictive contextual features to reduce uncertainty in random cost coefficients and thereby improve average-cost performance. An example is the stochastic shortest path problem with random edge costs (e.g., traffic) and contextual features (e.g., lagged traffic, weather). Existing work on CLO assumes the data has fully observed cost coefficient vectors, but in many applications, we can only see the realized cost of a historical decision, that is, just one projection of the random cost coefficient vector, to which we refer as bandit feedback. We study a class of offline learning algorithms for CLO with bandit feedback, which we term induced empirical risk minimization (IERM), where we fit a predictive model to directly optimize the downstream performance of the policy it induces. We show a fast-rate regret bound for IERM that allows for misspecified model classes and flexible choices of the optimization estimate, and we develop computationally tractable surrogate losses. A byproduct of our theory of independent interest is fast-rate regret bound for IERM with full feedback and misspecified policy class. We compare the performance of different modeling choices numerically using a stochastic shortest path example and provide practical insights from the empirical results.

Bayesian Optimization (BO) has proven to be very successful at optimizing a static, noisy, costly-to-evaluate black-box function $f : \mathcal{S} \to \mathbb{R}$. However, optimizing a black-box which is also a function of time (*i.e.*, a *dynamic* function) $f : \mathcal{S} \times \mathcal{T} \to \mathbb{R}$ remains a challenge, since a dynamic Bayesian Optimization (DBO) algorithm has to keep track of the optimum over time. This changes the nature of the optimization problem in at least three aspects: (i) querying an arbitrary point in $\mathcal{S} \times \mathcal{T}$ is impossible, (ii) past observations become less and less relevant for keeping track of the optimum as time goes by and (iii) the DBO algorithm must have a high sampling frequency so it can collect enough relevant observations to keep track of the optimum through time. In this paper, we design a Wasserstein distance-based criterion able to quantify the relevancy of an observation with respect to future predictions. Then, we leverage this criterion to build W-DBO, a DBO algorithm able to remove irrelevant observations from its dataset on the fly, thus maintaining simultaneously a good predictive performance and a high sampling frequency, even in continuous-time optimization tasks with unknown horizon. Numerical experiments establish the superiority of W-DBO, which outperforms state-of-the-art methods by a comfortable margin.

Hyperparameter Optimization (HPO) plays a pivotal role in unleashing the potential of iterative machine learning models. This paper addresses a crucial aspect that has largely been overlooked in HPO: the impact of uncertainty in ML model training. The paper introduces the concept of uncertainty-aware HPO and presents a novel approach called the UQ-guided scheme for quantifying uncertainty. This scheme offers a principled and versatile method to empower HPO techniques in handling model uncertainty during their exploration of the candidate space. By constructing a probabilistic model and implementing probability-driven candidate selection and budget allocation, this approach enhances the quality of the resulting model hyperparameters. It achieves a notable performance improvement of over 50\% in terms of accuracy regret and exploration time.

Neural Architecture Search (NAS) methods seek effective optimization toward performance metrics regarding model accuracy and generalization while facing challenges regarding search costs and GPU resources. Recent Neural Tangent Kernel (NTK) NAS methods achieve remarkable search efficiency based on a training-free model estimate; however, they overlook the non-convex nature of the DNNs in the search process. In this paper, we develop Multi-Objective Training-based Estimate (MOTE) for efficient NAS, retaining search effectiveness and achieving the new state-of-the-art in the accuracy and cost trade-off. To improve NTK and inspired by the Training Speed Estimation (TSE) method, MOTE is designed to model the actual performance of DNNs from macro to micro perspective by draw loss landscape and convergence speed simultaneously. Using two reduction strategies, the MOTE is generated based on a reduced architecture and a reduced dataset. Inspired by evolutionary search, our iterative ranking-based, coarse-to-fine architecture search is highly effective. Experiments on NASBench-201 show MOTE-NAS achieves 94.32% accuracy on CIFAR-10, 72.81% on CIFAR-100, and 46.38% on ImageNet-16-120, outperforming NTK-based NAS approaches. An evaluation-free (EF) version of MOTE-NAS delivers high efficiency in only 5 minutes, delivering a model more accurate than KNAS.

In-context learning (ICL) refers to a remarkable capability of pretrained large language models, which can learn a new task given a few examples during inference. However, theoretical understanding of ICL is largely under-explored, particularly whether transformers can be trained to generalize to unseen examples in a prompt, which will require the model to acquire contextual knowledge of the prompt for generalization. This paper investigates the training dynamics of transformers by gradient descent through the lens of non-linear regression tasks. The contextual generalization here can be attained via learning the template function for each task in-context, where all template functions lie in a linear space with $m$ basis functions. We analyze the training dynamics of one-layer multi-head transformers to {in-contextly} predict unlabeled inputs given partially labeled prompts, where the labels contain Gaussian noise and the number of examples in each prompt are not sufficient to determine the template. Under mild assumptions, we show that the training loss for a one-layer multi-head transformer converges linearly to a global minimum. Moreover, the transformer effectively learns to perform ridge regression over the basis functions. To our knowledge, this study is the first provable demonstration that transformers can learn contextual (i.e., template) information to generalize to both unseen examples and tasks when prompts contain only a small number of query-answer pairs.

Global polynomial optimization is an important tool across applied mathematics, with many applications in operations research, engineering, and the physical sciences. In various settings, the polynomials depend on external parameters that may be random. We discuss a stochastic sum-of-squares (S-SOS) algorithm based on the sum-of-squares hierarchy that constructs a series of semidefinite programs to jointly find strict lower bounds on the global minimum and extracts candidates for parameterized global minimizers. We prove quantitative convergence of the hierarchy as the degree increases and use it to solve unconstrained and constrained polynomial optimization problems parameterized by random variables. By employing n-body priors from condensed matter physics to induce sparsity, we can use S-SOS to produce solutions and uncertainty intervals for sensor network localization problems containing up to 40 variables and semidefinite matrix sizes surpassing 800 x 800.

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>.

The efficacy of large language models (LLMs) in understanding and generating natural language has aroused a wide interest in developing prompt-based methods to harness the power of black-box LLMs. Existing methodologies usually prioritize a global optimization for finding the global optimum, which however will perform poorly in certain tasks. This thus motivates us to re-think the necessity of finding a global optimum in prompt optimization. To answer this, we conduct a thorough empirical study on prompt optimization and draw two major insights. Contrasting with the rarity of global optimum, local optima are usually prevalent and well-performed, which can be more worthwhile for efficient prompt optimization (**Insight I**). The choice of the input domain, covering both the generation and the representation of prompts, affects the identification of well-performing local optima (**Insight II**). Inspired by these insights, we propose a novel algorithm, namely localized zeroth-order prompt optimization (ZOPO), which incorporates a Neural Tangent Kernel-based derived Gaussian process into standard zeroth-order optimization for an efficient search of well-performing local optima in prompt optimization. Remarkably, ZOPO outperforms existing baselines in terms of both the optimization performance and the query efficiency, which we demonstrate through extensive experiments.

Modern machine learning heavily depends on the effectiveness of optimization techniques. While deep learning models have achieved remarkable empirical results in training, their theoretical underpinnings remain somewhat elusive. Ensuring the convergence of optimization methods requires imposing specific structures on the objective function which often do not hold in practice. One prominent example is the widely recognized Polyak-Lojasiewicz (PL) inequality, which has garnered considerable attention in recent years. However, validating such assumptions for deep neural networks entails substantial and often impractical levels of over-parametrization. In order to address this limitation, we propose a novel class of functions that can characterize the loss landscape of modern deep models without requiring extensive over-parametrization and can also include saddle points. Crucially, we prove that gradient-based optimizers possess theoretical guarantees of convergence under this assumption. Finally, we validate the soundness of our assumption through both theoretical analysis and empirical experimentation across a diverse range of deep learning models.