NeurIPS 2023 论文列表

Learning to control unknown nonlinear dynamical systems is a fundamental problem in reinforcement learning and control theory. A commonly applied approach is to first explore the environment (exploration), learn an accurate model of it (system identification), and then compute an optimal controller with the minimum cost on this estimated system (policy optimization). While existing work has shown that it is possible to learn a uniformly good model of the system (Mania et al., 2020), in practice, if we aim to learn a good controller with a low cost on the actual system, certain system parameters may be significantly more critical than others, and we therefore ought to focus our exploration on learning such parameters. In this work, we consider the setting of nonlinear dynamical systems and seek to formally quantify, in such settings, (a) which parameters are most relevant to learning a good controller, and (b) how we can best explore so as to minimize uncertainty in such parameters. Inspired by recent work in linear systems (Wagenmaker et al., 2021), we show that minimizing the controller loss in nonlinear systems translates to estimating the system parameters in a particular, task-dependent metric. Motivated by this, we develop an algorithm able to efficiently explore the system to reduce uncertainty in this metric, and prove a lower bound showing that our approach learns a controller at a near-instance-optimal rate. Our algorithm relies on a general reduction from policy optimization to optimal experiment design in arbitrary systems, and may be of independent interest. We conclude with experiments demonstrating the effectiveness of our method in realistic nonlinear robotic systems.

Large vision models with billions of parameters and trained on broad data have great potential in numerous downstream applications. However, these models are typically difficult to adapt due to their large parameter size and sometimes lack of accesss to their weights---entities able to develop large vision models often provide APIs only. In this paper, we study how to better utilize large vision models through the lens of in-context learning, a concept that has been well-known in natural language processing but has only been studied very recently in computer vision. In-context learning refers to the ability to perform inference on tasks never seen during training by simply conditioning on in-context examples (i.e., input-output pairs) without updating any internal model parameters. To demystify in-context learning in computer vision, we conduct an extensive research and identify a critical problem: downstream performance is highly sensitivie to the choice of visual in-context examples. To address this problem, we propose a prompt retrieval framework specifically for large vision models, allowing the selection of in-context examples to be fully automated. Concretely, we provide two implementations: (i) an unsupervised prompt retrieval method based on nearest example search using an off-the-shelf model, and (ii) a supervised prompt retrieval method, which trains a neural network to choose examples that directly maximize in-context learning performance. Both methods do not require access to the internal weights of large vision models. Our results demonstrate that our methods can bring non-trivial improvements to visual in-context learning in comparison to the commonly-used random selection. Code and models will be released.

Causal identification is at the core of the causal inference literature, where complete algorithms have been proposed to identify causal queries of interest. The validity of these algorithms hinges on the restrictive assumption of having access to a correctly specified causal structure. In this work, we study the setting where a probabilistic model of the causal structure is available. Specifically, the edges in a causal graph exist with uncertainties which may, for example, represent degree of belief from domain experts. Alternatively, the uncertainty about an edge may reflect the confidence of a particular statistical test. The question that naturally arises in this setting is: Given such a probabilistic graph and a specific causal effect of interest, what is the subgraph which has the highest plausibility and for which the causal effect is identifiable? We show that answering this question reduces to solving an NP-hard combinatorial optimization problem which we call the edge ID problem. We propose efficient algorithms to approximate this problem and evaluate them against both real-world networks and randomly generated graphs.

Semi-Supervised Learning (SSL) has been an effective way to leverage abundant unlabeled data with extremely scarce labeled data. However, most SSL methods are commonly based on instance-wise consistency between different data transformations. Therefore, the label guidance on labeled data is hard to be propagated to unlabeled data. Consequently, the learning process on labeled data is much faster than on unlabeled data which is likely to fall into a local minima that does not favor unlabeled data, leading to sub-optimal generalization performance. In this paper, we propose FlatMatch which minimizes a cross-sharpness measure to ensure consistent learning performance between the two datasets. Specifically, we increase the empirical risk on labeled data to obtain a worst-case model which is a failure case needing to be enhanced. Then, by leveraging the richness of unlabeled data, we penalize the prediction difference (i.e., cross-sharpness) between the worst-case model and the original model so that the learning direction is beneficial to generalization on unlabeled data. Therefore, we can calibrate the learning process without being limited to insufficient label information. As a result, the mismatched learning performance can be mitigated, further enabling the effective exploitation of unlabeled data and improving SSL performance. Through comprehensive validation, we show FlatMatch achieves state-of-the-art results in many SSL settings.

While different neural models often exhibit latent spaces that are alike when exposed to semantically related data, this intrinsic similarity is not always immediately discernible. Towards a better understanding of this phenomenon, our work shows how representations learned from these neural modules can be translated between different pre-trained networks via simpler transformations than previously thought. An advantage of this approach is the ability to estimate these transformations using standard, well-understood algebraic procedures that have closed-form solutions. Our method directly estimates a transformation between two given latent spaces, thereby enabling effective stitching of encoders and decoders without additional training. We extensively validate the adaptability of this translation procedure in different experimental settings: across various trainings, domains, architectures (e.g., ResNet, CNN, ViT), and in multiple downstream tasks (classification, reconstruction). Notably, we show how it is possible to zero-shot stitch text encoders and vision decoders, or vice-versa, yielding surprisingly good classification performance in this multimodal setting.

Classification is a fundamental task in science and engineering on which machine learning methods have shown outstanding performances. However, it is challenging to determine whether such methods have achieved the Bayes error rate, that is, the lowest error rate attained by any classifier. This is mainly due to the fact that the Bayes error rate is not known in general and hence, effectively estimating it is paramount. Inspired by the work by Ishida et al. (2023), we propose an estimator for the Bayes error rate of supervised multi-class classification problems. We analyze several theoretical aspects of such estimator, including its consistency, unbiasedness, convergence rate, variance, and robustness. We also propose a denoising method that reduces the noise that potentially corrupts the data labels, and we improve the robustness of the proposed estimator to outliers by incorporating the median-of-means estimator. Our analysis demonstrates the consistency, asymptotic unbiasedness, convergence rate, and robustness of the proposed estimators. Finally, we validate the effectiveness of our theoretical results via experiments both on synthetic data under various noise settings and on real data.

Finding the proper depth $d$ of a graph convolutional network (GCN) that provides strong representation ability has drawn significant attention, yet nonetheless largely remains an open problem for the graph learning community. Although noteworthy progress has been made, the depth or the number of layers of a corresponding GCN is realized by a series of graph convolution operations, which naturally makes $d$ a positive integer ($d \in \mathbb{N}+$). An interesting question is whether breaking the constraint of $\mathbb{N}+$ by making $d$ a real number ($d \in \mathbb{R}$) can bring new insights into graph learning mechanisms. In this work, by redefining GCN's depth $d$ as a trainable parameter continuously adjustable within $(-\infty,+\infty)$, we open a new door of controlling its signal processing capability to model graph homophily/heterophily (nodes with similar/dissimilar labels/attributes tend to be inter-connected). A simple and powerful GCN model TEDGCN, is proposed to retain the simplicity of GCN and meanwhile automatically search for the optimal $d$ without the prior knowledge regarding whether the input graph is homophilic or heterophilic. Negative-valued $d$ intrinsically enables high-pass frequency filtering functionality via augmented topology for graph heterophily. Extensive experiments demonstrate the superiority of TEDGCN on node classification tasks for a variety of homophilic and heterophilic graphs.

Audio-visual learning has been a major pillar of multi-modal machine learning, where the community mostly focused on its $\textit{modality-aligned}$ setting, $\textit{i.e.}$, the audio and visual modality are $\textit{both}$ assumed to signal the prediction target. With the Look, Listen, and Parse dataset (LLP), we investigate the under-explored $\textit{unaligned}$ setting, where the goal is to recognize audio and visual events in a video with only weak labels observed. Such weak video-level labels only tell what events happen without knowing the modality they are perceived (audio, visual, or both). To enhance learning in this challenging setting, we incorporate large-scale contrastively pre-trained models as the modality teachers. A simple, effective, and generic method, termed $\textbf{V}$isual-$\textbf{A}$udio $\textbf{L}$abel Elab$\textbf{or}$ation (VALOR), is innovated to harvest modality labels for the training events. Empirical studies show that the harvested labels significantly improve an attentional baseline by $\textbf{8.0}$ in average F-score (Type@AV). Surprisingly, we found that modality-independent teachers outperform their modality-fused counterparts since they are noise-proof from the other potentially unaligned modality. Moreover, our best model achieves the new state-of-the-art on all metrics of LLP by a substantial margin ($\textbf{+5.4}$ F-score for Type@AV). VALOR is further generalized to Audio-Visual Event Localization and achieves the new state-of-the-art as well.

There is currently a significant gap between the performance of fine-tuned models and prompting approaches using Large Language Models (LLMs) on the challenging task of text-to-SQL, as evaluated on datasets such as Spider. To improve the performance of LLMs in the reasoning process, we study how decomposing the task into smaller sub-tasks can be effective. In particular, we show that breaking down the generation problem into sub-problems and feeding the solutions of those sub-problems into LLMs can be an effective approach for significantly improving their performance. Our experiments with three LLMs show that this approach consistently improves their simple few-shot performance by roughly 10%, pushing the accuracy of LLMs towards SOTA or surpassing it. On the holdout test set of Spider, the SOTA, in terms of execution accuracy, was 79.9 and the new SOTA at the time of this writing using our approach is 85.3. Our approach with in-context learning beats many heavily fine-tuned models by at least 5%. Additionally, when evaluated on the BIRD benchmark, our approach achieved an execution accuracy of 55.9%, setting a new SOTA on its holdout test set.

Many modern machine learning applications - from online principal component analysis to covariance matrix identification and dictionary learning - can be formulated as minimization problems on Riemannian manifolds, typically solved with a Riemannian stochastic gradient method (or some variant thereof). However, in many cases of interest, the resulting minimization problem is _not_ geodesically convex, so the convergence of the chosen solver to a desirable solution - i.e., a local minimizer - is by no means guaranteed. In this paper, we study precisely this question, that is, whether stochastic Riemannian optimization algorithms are guaranteed to avoid saddle points with probability $1$. For generality, we study a family of retraction-based methods which, in addition to having a potentially much lower per-iteration cost relative to Riemannian gradient descent, include other widely used algorithms, such as natural policy gradient methods and mirror descent in ordinary convex spaces. In this general setting, we show that, under mild assumptions for the ambient manifold and the oracle providing gradient information, the policies under study avoid strict saddle points / submanifolds with probability $1$, from any initial condition. This result provides an important sanity check for the use of gradient methods on manifolds as it shows that, almost always, the end state of a stochastic Riemannian algorithm can only be a local minimizer.

Pre-trained language models can be surprisingly adept at tasks they were not explicitly trained on, but how they implement these capabilities is poorly understood. In this paper, we investigate the basic mathematical abilities often acquired by pre-trained language models. Concretely, we use mechanistic interpretability techniques to explain the (limited) mathematical abilities of GPT-2 small. As a case study, we examine its ability to take in sentences such as "The war lasted from the year 1732 to the year 17", and predict valid two-digit end years (years > 32). We first identify a circuit, a small subset of GPT-2 small's computational graph that computes this task's output. Then, we explain the role of each circuit component, showing that GPT-2 small's final multi-layer perceptrons boost the probability of end years greater than the start year. Finally, we find related tasks that activate our circuit. Our results suggest that GPT-2 small computes greater-than using a complex but general mechanism that activates across diverse contexts.

Recent AI-assistant agents, such as ChatGPT, predominantly rely on supervised fine-tuning (SFT) with human annotations and reinforcement learning from human feedback (RLHF) to align the output of large language models (LLMs) with human intentions, ensuring they are helpful, ethical, and reliable. However, this dependence can significantly constrain the true potential of AI-assistant agents due to the high cost of obtaining human supervision and the related issues on quality, reliability, diversity, self-consistency, and undesirable biases. To address these challenges, we propose a novel approach called SELF-ALIGN, which combines principle-driven reasoning and the generative power of LLMs for the self-alignment of AI agents with minimal human supervision. Our approach encompasses four stages: first, we use an LLM to generate synthetic prompts, and a topic-guided method to augment the prompt diversity; second, we use a small set of human-written principles for AI models to follow, and guide the LLM through in-context learning from demonstrations (of principles application) to produce helpful, ethical, and reliable responses to user's queries; third, we fine-tune the original LLM with the high-quality self-aligned responses so that the resulting model can generate desirable responses for each query directly without the principle set and the demonstrations anymore; and finally, we offer a refinement step to address the issues of overly-brief or indirect responses. Applying SELF-ALIGN to the LLaMA-65b base language model, we develop an AI assistant named Dromedary. With fewer than 300 lines of human annotations (including < 200 seed prompts, 16 generic principles, and 5 exemplars for in-context learning). Dromedary significantly surpasses the performance of several state-of-the-art AI systems, including Text-Davinci-003 and Alpaca, on benchmark datasets with various settings.

Action-constrained reinforcement learning (ACRL) is a popular approach for solving safety-critical and resource-allocation related decision making problems. A major challenge in ACRL is to ensure agent taking a valid action satisfying constraints in each RL step. Commonly used approach of using a projection layer on top of the policy network requires solving an optimization program which can result in longer training time, slow convergence, and zero gradient problem. To address this, first we use a normalizing flow model to learn an invertible, differentiable mapping between the feasible action space and the support of a simple distribution on a latent variable, such as Gaussian. Second, learning the flow model requires sampling from the feasible action space, which is also challenging. We develop multiple methods, based on Hamiltonian Monte-Carlo and probabilistic sentential decision diagrams for such action sampling for convex and non-convex constraints. Third, we integrate the learned normalizing flow with the DDPG algorithm. By design, a well-trained normalizing flow will transform policy output into a valid action without requiring an optimization solver. Empirically, our approach results in significantly fewer constraint violations (upto an order-of-magnitude for several instances) and is multiple times faster on a variety of continuous control tasks.

Urban mobility analysis has been extensively studied in the past decade using a vast amount of GPS trajectory data, which reveals hidden patterns in movement and human activity within urban landscapes. Despite its significant value, the availability of such datasets often faces limitations due to privacy concerns, proprietary barriers, and quality inconsistencies. To address these challenges, this paper presents a synthetic trajectory dataset with high fidelity, offering a general solution to these data accessibility issues. Specifically, the proposed dataset adopts a diffusion model as its synthesizer, with the primary aim of accurately emulating the spatial-temporal behavior of the original trajectory data. These synthesized data can retain the geo-distribution and statistical properties characteristic of real-world datasets. Through rigorous analysis and case studies, we validate the high similarity and utility between the proposed synthetic trajectory dataset and real-world counterparts. Such validation underscores the practicality of synthetic datasets for urban mobility analysis and advocates for its wider acceptance within the research community. Finally, we publicly release the trajectory synthesizer and datasets, aiming to enhance the quality and availability of synthetic trajectory datasets and encourage continued contributions to this rapidly evolving field. The dataset is released for public online availability https://github.com/Applied-Machine-Learning-Lab/SynMob.

Regularization in modern machine learning is crucial, and it can take various forms in algorithmic design: training set, model family, error function, regularization terms, and optimizations. In particular, the learning rate, which can be interpreted as a temperature-like parameter within the statistical mechanics of learning, plays a crucial role in neural network training. Indeed, many widely adopted training strategies basically just define the decay of the learning rate over time. This process can be interpreted as decreasing a temperature, using either a global learning rate (for the entire model) or a learning rate that varies for each parameter. This paper proposes TempBalance, a straightforward yet effective layer-wise learning rate method. TempBalance is based on Heavy-Tailed Self-Regularization (HT-SR) Theory, an approach which characterizes the implicit self-regularization of different layers in trained models. We demonstrate the efficacy of using HT-SR-motivated metrics to guide the scheduling and balancing of temperature across all network layers during model training, resulting in improved performance during testing. We implement TempBalance on CIFAR10, CIFAR100, SVHN, and TinyImageNet datasets using ResNets, VGGs and WideResNets with various depths and widths. Our results show that TempBalance significantly outperforms ordinary SGD and carefully-tuned spectral norm regularization. We also show that TempBalance outperforms a number of state-of-the-art optimizers and learning rate schedulers.

Recent contrastive language image pre-training has led to learning highly transferable and robust image representations. However, adapting these models to video domain with minimal supervision remains an open problem. We explore a simple step in that direction, using language tied self-supervised learning to adapt an image CLIP model to the video domain. A backbone modified for temporal modeling is trained under self-distillation settings with train objectives operating in an action concept space. Feature vectors of various action concepts extracted from a language encoder using relevant textual prompts construct this space. A large language model aware of actions and their attributes generates the relevant textual prompts. We introduce two train objectives, concept distillation and concept alignment, that retain generality of original representations while enforcing relations between actions and their attributes. Our approach improves zero-shot and linear probing performance on three action recognition benchmarks.

With the advent of deep learning, progressively larger neural networks have been designed to solve complex tasks. We take advantage of these capacity-rich models to lower the cost of inference by exploiting computation in superposition. To reduce the computational burden per input, we propose Multiple-Input-Multiple-Output Neural Networks (MIMONets) capable of handling many inputs at once. MIMONets augment various deep neural network architectures with variable binding mechanisms to represent an arbitrary number of inputs in a compositional data structure via fixed-width distributed representations. Accordingly, MIMONets adapt nonlinear neural transformations to process the data structure holistically, leading to a speedup nearly proportional to the number of superposed input items in the data structure. After processing in superposition, an unbinding mechanism recovers each transformed input of interest. MIMONets also provide a dynamic trade-off between accuracy and throughput by an instantaneous on-demand switching between a set of accuracy-throughput operating points, yet within a single set of fixed parameters. We apply the concept of MIMONets to both CNN and Transformer architectures resulting in MIMOConv and MIMOFormer, respectively. Empirical evaluations show that MIMOConv achieves $\approx 2$–$4\times$ speedup at an accuracy delta within [+0.68, -3.18]% compared to WideResNet CNNs on CIFAR10 and CIFAR100. Similarly, MIMOFormer can handle $2$–$4$ inputs at once while maintaining a high average accuracy within a [-1.07, -3.43]% delta on the long range arena benchmark. Finally, we provide mathematical bounds on the interference between superposition channels in MIMOFormer. Our code is available at https://github.com/IBM/multiple-input-multiple-output-nets.

Kullback-Leibler (KL) divergence is one of the most important measures to calculate the difference between probability distributions. In this paper, we theoretically study several properties of KL divergence between multivariate Gaussian distributions. Firstly, for any two $n$-dimensional Gaussian distributions $\mathcal{N}_1$ and $\mathcal{N}_2$, we prove that when $KL(\mathcal{N}_2||\mathcal{N}_1)\leq \varepsilon\ (\varepsilon>0)$ the supremum of $KL(\mathcal{N}_1||\mathcal{N}_2)$ is $(1/2)\left((-W_{0}(-e^{-(1+2\varepsilon)}))^{-1}+\log(-W_{0}(-e^{-(1+2\varepsilon)})) -1 \right)$, where $W_0$ is the principal branch of Lambert $W$ function. For small $\varepsilon$, the supremum is $\varepsilon + 2\varepsilon^{1.5} + O(\varepsilon^2)$. This quantifies the approximate symmetry of small KL divergence between Gaussian distributions. We further derive the infimum of $KL(\mathcal{N}_1||\mathcal{N}_2)$ when $KL(\mathcal{N}_2||\mathcal{N}_1)\geq M\ (M>0)$. We give the conditions when the supremum and infimum can be attained. Secondly, for any three $n$-dimensional Gaussian distributions $\mathcal{N}_1$, $\mathcal{N}_2$, and $\mathcal{N}_3$, we theoretically show that an upper bound of $KL(\mathcal{N}_1||\mathcal{N}_3)$ is $3\varepsilon_1+3\varepsilon_2+2\sqrt{\varepsilon_1\varepsilon_2}+o(\varepsilon_1)+o(\varepsilon_2)$ when $KL(\mathcal{N}_1||\mathcal{N}_2)\leq \varepsilon_1$ and $KL(\mathcal{N}_2||\mathcal{N}_3)\leq \varepsilon_2$ ($\varepsilon_1,\varepsilon_2\ge 0$). This reveals that KL divergence between Gaussian distributions follows a relaxed triangle inequality. Note that, all these bounds in the theorems presented in this work are independent of the dimension $n$. Finally, we discuss several applications of our theories in deep learning, reinforcement learning, and sample complexity research.

The local search methods have been widely used to solve the clustering problems. In practice, local search algorithms for clustering problems mainly adapt the single-swap strategy, which enables them to handle large-scale datasets and achieve linear running time in the data size. However, compared with multi-swap local search algorithms, there is a considerable gap on the approximation ratios of the single-swap local search algorithms. Although the current multi-swap local search algorithms provide small constant approximation, the proposed algorithms tend to have large polynomial running time, which cannot be used to handle large-scale datasets. In this paper, we propose a multi-swap local search algorithm for the $k$-means problem with linear running time in the data size. Given a swap size $t$, our proposed algorithm can achieve a $(50(1+\frac{1}{t})+\epsilon)$-approximation, which improves the current best result 509 (ICML 2019) with linear running time in the data size. Our proposed method, compared with previous multi-swap local search algorithms, is the first one to achieve linear running time in the data size. To obtain a more practical algorithm for the problem with better clustering quality and running time, we propose a sampling-based method which accelerates the process of clustering cost update during swaps. Besides, a recombination mechanism is proposed to find potentially better solutions. Empirical experiments show that our proposed algorithms achieve better performances compared with branch and bound solver (NeurIPS 2022) and other existing state-of-the-art local search algorithms on both small and large datasets.