NeurIPS 2023 论文列表

Computational tractability and social welfare (aka. efficiency) of equilibria are two fundamental but in general orthogonal considerations in algorithmic game theory. Nevertheless, we show that when (approximate) full efficiency can be guaranteed via a smoothness argument a la Roughgarden, Nash equilibria are approachable under a family of no-regret learning algorithms, thereby enabling fast and decentralized computation. We leverage this connection to obtain new convergence results in large games---wherein the number of players $n \gg 1$---under the well-documented property of full efficiency via smoothness in the limit. Surprisingly, our framework unifies equilibrium computation in disparate classes of problems including games with vanishing strategic sensitivity and two-player zero-sum games, illuminating en route an immediate but overlooked equivalence between smoothness and a well-studied condition in the optimization literature known as the Minty property. Finally, we establish that a family of no-regret dynamics attains a welfare bound that improves over the smoothness framework while at the same time guaranteeing convergence to the set of coarse correlated equilibria. We show this by employing the clairvoyant mirror descent algortihm recently introduced by Piliouras et al.

Transformer architecture has shown impressive performance in multiple research domains and has become the backbone of many neural network models. However, there is limited understanding on how it works. In particular, with a simple predictive loss, how the representation emerges from the gradient \emph{training dynamics} remains a mystery. In this paper, for 1-layer transformer with one self-attention layer plus one decoder layer, we analyze its SGD training dynamics for the task of next token prediction in a mathematically rigorous manner. We open the black box of the dynamic process of how the self-attention layer combines input tokens, and reveal the nature of underlying inductive bias. More specifically, with the assumption (a) no positional encoding, (b) long input sequence, and (c) the decoder layer learns faster than the self-attention layer, we prove that self-attention acts as a \emph{discriminative scanning algorithm}: starting from uniform attention, it gradually attends more to distinct key tokens for a specific next token to be predicted, and pays less attention to common key tokens that occur across different next tokens. Among distinct tokens, it progressively drops attention weights, following the order of low to high co-occurrence between the key and the query token in the training set. Interestingly, this procedure does not lead to winner-takes-all, but stops due to a \emph{phase transition} that is controllable by the learning rate of the decoder layer, leaving (almost) fixed token combination. We verify this \textbf{\emph{scan and snap}} dynamics on synthetic and real-world data (WikiText-103).

Learning controllers with offline data in decision-making systems is an essential area of research due to its potential to reduce the risk of applications in real-world systems. However, in responsibility-sensitive settings such as healthcare, decision accountability is of paramount importance, yet has not been adequately addressed by the literature. This paper introduces the Accountable Offline Controller (AOC) that employs the offline dataset as the Decision Corpus and performs accountable control based on a tailored selection of examples, referred to as the Corpus Subset. AOC operates effectively in low-data scenarios, can be extended to the strictly offline imitation setting, and displays qualities of both conservation and adaptability. We assess AOC's performance in both simulated and real-world healthcare scenarios, emphasizing its capability to manage offline control tasks with high levels of performance while maintaining accountability.

We consider offline reinforcement learning (RL) where we only have only access to offline data. In contrast to numerous offline RL algorithms that necessitate the uniform coverage of the offline data over state and action space, we propose value-based algorithms with PAC guarantees under partial coverage, specifically, coverage of offline data against a single policy, and realizability of soft Q-function (a.k.a., entropy-regularized Q-function) and another function, which is defined as a solution to a saddle point of certain minimax optimization problem). Furthermore, we show the analogous result for Q-functions instead of soft Q-functions. To attain these guarantees, we use novel algorithms with minimax loss functions to accurately estimate soft Q-functions and Q-functions with -convergence guarantees measured on the offline data. We introduce these loss functions by casting the estimation problems into nonlinear convex optimization problems and taking the Lagrange functions.

Recent work has shown that generation from a prompted or fine-tuned language model can perform well at semantic parsing when the output is constrained to be a valid semantic representation. We introduce BenchCLAMP, a Benchmark to evaluate Constrained LAnguage Model Parsing, that includes context-free grammars for seven semantic parsing datasets and two syntactic parsing datasets with varied output meaning representations, as well as a constrained decoding interface to generate only valid outputs covered by these grammars. We provide low, medium, and high resource splits for each dataset, allowing accurate comparison of various language models under different data regimes. Our benchmark supports evaluation of language models using prompt-based learning as well as fine-tuning. We benchmark seven language models, including two GPT-3 variants available only through an API. Our experiments show that encoder-decoder pretrained language models can achieve similar performance or even surpass state-of-the-art methods for both syntactic and semantic parsing when the model output is constrained to be valid.

The increasing use of neural networks in various applications has lead to increasing apprehensions, underscoring the necessity to understand their operations beyond mere final predictions. As a solution to enhance model transparency, Concept Bottleneck Models (CBMs) have gained popularity since their introduction. CBMs essentially limit the latent space of a model to human-understandable high-level concepts. While beneficial, CBMs have been reported to often learn irrelevant concept representations that consecutively damage model performance. To overcome the performance trade-off, we propose a cooperative-Concept Bottleneck Model (coop-CBM). The concept representation of our model is particularly meaningful when fine-grained concept labels are absent. Furthermore, we introduce the concept orthogonal loss (COL) to encourage the separation between the concept representations and to reduce the intra-concept distance. This paper presents extensive experiments on real-world datasets for image classification tasks, namely CUB, AwA2, CelebA and TIL. We also study the performance of coop-CBM models under various distributional shift settings. We show that our proposed method achieves higher accuracy in all distributional shift settings even compared to the black-box models with the highest concept accuracy.

We present a simple but effective method to measure and mitigate model biases caused by reliance on spurious cues. Instead of requiring costly changes to one's data or model training, our method better utilizes the data one already has by sorting them. Specifically, we rank images within their classes based on spuriosity (the degree to which common spurious cues are present), proxied via deep neural features of an interpretable network. With spuriosity rankings, it is easy to identify minority subpopulations (i.e. low spuriosity images) and assess model bias as the gap in accuracy between high and low spuriosity images. One can even efficiently remove a model's bias at little cost to accuracy by finetuning its classification head on low spuriosity images, resulting in fairer treatment of samples regardless of spuriosity. We demonstrate our method on ImageNet, annotating $5000$ class-feature dependencies ($630$ of which we find to be spurious) and generating a dataset of $325k$ soft segmentations for these features along the way. Having computed spuriosity rankings via the identified spurious neural features, we assess biases for $89$ diverse models and find that class-wise biases are highly correlated across models. Our results suggest that model bias due to spurious feature reliance is influenced far more by what the model is trained on than how it is trained.

Denoising diffusion probabilistic models have transformed image generation with their impressive fidelity and diversity. We show that they also excel in estimating optical flow and monocular depth, surprisingly without task-specific architectures and loss functions that are predominant for these tasks. Compared to the point estimates of conventional regression-based methods, diffusion models also enable Monte Carlo inference, e.g., capturing uncertainty and ambiguity in flow and depth. With self-supervised pre-training, the combined use of synthetic and real data for supervised training, and technical innovations (infilling and step-unrolled denoising diffusion training) to handle noisy-incomplete training data, one can train state-of-the-art diffusion models for depth and optical flow estimation, with additional zero-shot coarse-to-fine refinement for high resolution estimates. Extensive experiments focus on quantitative performance against benchmarks, ablations, and the model's ability to capture uncertainty and multimodality, and impute missing values. Our model obtains a state-of-the-art relative depth error of 0.074 on the indoor NYU benchmark and an Fl-all score of 3.26\% on the KITTI optical flow benchmark, about 25\% better than the best published method.

Masked pre-training removes random input dimensions and learns a model that can predict the missing values. Empirical results indicate that this intuitive form of self-supervised learning yields models that generalize very well to new domains. A theoretical understanding is, however, lacking. This paper shows that masked pre-training with a suitable cumulative scoring function corresponds to maximizing the model's marginal likelihood, which is de facto the Bayesian model selection measure of generalization. Beyond shedding light on the success of masked pre-training, this insight also suggests that Bayesian models can be trained with appropriately designed self-supervision. Empirically, we confirm the developed theory and explore the main learning principles of masked pre-training in large language models.

Federated learning (FL) is a promising privacy-preserving machine learning paradigm over distributed data. In this paradigm, each client trains the parameter of a model locally and the server aggregates the parameter from clients periodically. Therefore, we perform the learning and communication over the same set of parameters. However, we find that learning and communication have fundamentally divergent requirements for parameter selection, akin to two opposite teams in a tug-of-war game. To mitigate this discrepancy, we introduce FedSep, a novel two-layer federated learning framework. FedSep consists of separated communication and learning layers for each client and the two layers are connected through decode/encode operations. In particular, the decoding operation is formulated as a minimization problem. We view FedSep as a federated bilevel optimization problem and propose an efficient algorithm to solve it. Theoretically, we demonstrate that its convergence matches that of the standard FL algorithms. The separation of communication and learning in FedSep offers innovative solutions to various challenging problems in FL, such as Communication-Efficient FL and Heterogeneous-Model FL. Empirical validation shows the superior performance of FedSep over various baselines in these tasks.

In this paper, we critically examine the prevalent practice of using additive mixtures of Mat\'ern kernels in single-output Gaussian process (GP) models and explore the properties of multiplicative mixtures of Mat\'ern kernels for multi-output GP models. For the single-output case, we derive a series of theoretical results showing that the smoothness of a mixture of Mat\'ern kernels is determined by the least smooth component and that a GP with such a kernel is effectively equivalent to the least smooth kernel component. Furthermore, we demonstrate that none of the mixing weights or parameters within individual kernel components are identifiable. We then turn our attention to multi-output GP models and analyze the identifiability of the covariance matrix $A$ in the multiplicative kernel $K(x,y) = AK_0(x,y)$, where $K_0$ is a standard single output kernel such as Mat\'ern. We show that $A$ is identifiable up to a multiplicative constant, suggesting that multiplicative mixtures are well suited for multi-output tasks. Our findings are supported by extensive simulations and real applications for both single- and multi-output settings. This work provides insight into kernel selection and interpretation for GP models, emphasizing the importance of choosing appropriate kernel structures for different tasks.

In voting theory, when voters have ranked preferences over candidates, the celebrated Gibbard-Satterthwaite Theorem essentially rules out the existence of reasonable strategyproof methods for picking a winner. What if we weaken strategyproofness to only hold for Bayesian voters with beliefs over others' preferences? When voters believe other participants' rankings are drawn independently from a fixed distribution, the impossibility persists. However, it is quite reasonable for a voter to believe that other votes are correlated, either to each other or to their own ranking. We consider such beliefs induced by classic probabilistic models in social choice such as the Mallows, Placket-Luce, and Thurstone-Mosteller models. We single out the plurality rule (choosing the candidate ranked first most often) as a particularly promising choice as it is strategyproof for a large class of beliefs containing the specific ones we introduce. Further, we show that plurality is unique among positional scoring rules in having this property: no other scoring rule is strategyproof for beliefs induced by the Mallows model when there are a sufficient number of voters. Finally, we give examples of prominent non-scoring voting rules failing to be strategyproof on beliefs in this class, further bolstering the case for plurality.

Several recent methods for interpretability model feature interactions by looking at the Hessian of a neural network. This poses a challenge for ReLU networks, which are piecewise-linear and thus have a zero Hessian almost everywhere. We propose SmoothHess, a method of estimating second-order interactions through Stein's Lemma. In particular, we estimate the Hessian of the network convolved with a Gaussian through an efficient sampling algorithm, requiring only network gradient calls. SmoothHess is applied post-hoc, requires no modifications to the ReLU network architecture, and the extent of smoothing can be controlled explicitly. We provide a non-asymptotic bound on the sample complexity of our estimation procedure. We validate the superior ability of SmoothHess to capture interactions on benchmark datasets and a real-world medical spirometry dataset.

In computational neuroscience, there has been an increased interest in developing machine learning algorithms that leverage brain imaging data to provide estimates of "brain age" for an individual. Importantly, the discordance between brain age and chronological age (referred to as "brain age gap") can capture accelerated aging due to adverse health conditions and therefore, can reflect increased vulnerability towards neurological disease or cognitive impairments. However, widespread adoption of brain age for clinical decision support has been hindered due to lack of transparency and methodological justifications in most existing brain age prediction algorithms. In this paper, we leverage coVariance neural networks (VNN) to propose an explanation-driven and anatomically interpretable framework for brain age prediction using cortical thickness features. Specifically, our brain age prediction framework extends beyond the coarse metric of brain age gap in Alzheimer’s disease (AD) and we make two important observations: (i) VNNs can assign anatomical interpretability to elevated brain age gap in AD by identifying contributing brain regions, (ii) the interpretability offered by VNNs is contingent on their ability to exploit specific eigenvectors of the anatomical covariance matrix. Together, these observations facilitate an explainable and anatomically interpretable perspective to the task of brain age prediction.

We study the problems of sequential nonparametric two-sample and independence testing. Sequential tests process data online and allow using observed data to decide whether to stop and reject the null hypothesis or to collect more data, while maintaining type I error control. We build upon the principle of (nonparametric) testing by betting, where a gambler places bets on future observations and their wealth measures evidence against the null hypothesis. While recently developed kernel-based betting strategies often work well on simple distributions, selecting a suitable kernel for high-dimensional or structured data, such as images, is often nontrivial. To address this drawback, we design prediction-based betting strategies that rely on the following fact: if a sequentially updated predictor starts to consistently determine (a) which distribution an instance is drawn from, or (b) whether an instance is drawn from the joint distribution or the product of the marginal distributions (the latter produced by external randomization), it provides evidence against the two-sample or independence nulls respectively. We empirically demonstrate the superiority of our tests over kernel-based approaches under structured settings. Our tests can be applied beyond the case of independent and identically distributed data, remaining valid and powerful even when the data distribution drifts over time.

Concept erasure aims to remove specified features from a representation. It can improve fairness (e.g. preventing a classifier from using gender or race) and interpretability (e.g. removing a concept to observe changes in model behavior). We introduce LEAst-squares Concept Erasure (LEACE), a closed-form method which provably prevents all linear classifiers from detecting a concept while changing the representation as little as possible, as measured by a broad class of norms. We apply LEACE to large language models with a novel procedure called concept scrubbing, which erases target concept information from _every_ layer in the network. We demonstrate our method on two tasks: measuring the reliance of language models on part-of-speech information, and reducing gender bias in BERT embeddings. Our code is available at https://github.com/EleutherAI/concept-erasure.

In computational social choice, the distortion of a voting rule quantifies the degree to which the rule overcomes limited preference information to select a socially desirable outcome. This concept has been investigated extensively, but only through a worst-case lens. Instead, we study the expected distortion of voting rules with respect to an underlying distribution over voter utilities. Our main contribution is the design and analysis of a novel and intuitive rule, binomial voting, which provides strong distribution-independent guarantees for both expected distortion and expected welfare.

Spectral clustering is a popular and effective algorithm designed to find $k$ clusters in a graph $G$. In the classical spectral clustering algorithm, the vertices of $G$ are embedded into $\mathbb{R}^k$ using $k$ eigenvectors of the graph Laplacian matrix. However, computing this embedding is computationally expensive and dominates the running time of the algorithm. In this paper, we present a simple spectral clustering algorithm based on a vertex embedding with $O(\log(k))$ vectors computed by the power method. The vertex embedding is computed in nearly-linear time with respect to the size of the graph, and the algorithm provably recovers the ground truth clusters under natural assumptions on the input graph. We evaluate the new algorithm on several synthetic and real-world datasets, finding that it is significantly faster than alternative clustering algorithms, while producing results with approximately the same clustering accuracy.

Transformer interpretability aims to understand the algorithm implemented by a learned Transformer by examining various aspects of the model, such as the weight matrices or the attention patterns. In this work, through a combination of theoretical results and carefully controlled experiments on synthetic data, we take a critical view of methods that exclusively focus on individual parts of the model, rather than consider the network as a whole. We consider a simple synthetic setup of learning a (bounded) Dyck language. Theoretically, we show that the set of models that (exactly or approximately) solve this task satisfy a structural characterization derived from ideas in formal languages (the pumping lemma). We use this characterization to show that the set of optima is qualitatively rich; in particular, the attention pattern of a single layer can be "nearly randomized", while preserving the functionality of the network. We also show via extensive experiments that these constructions are not merely a theoretical artifact: even with severe constraints to the architecture of the model, vastly different solutions can be reached via standard training. Thus, interpretability claims based on inspecting individual heads or weight matrices in the Transformer can be misleading.