NeurIPS 2023 Papers

One of the fundamental challenges found throughout the data sciences is to explain why things happen in specific ways, or through which mechanisms a certain variable $X$ exerts influences over another variable $Y$. In statistics and machine learning, significant efforts have been put into developing machinery to estimate correlations across variables efficiently. In causal inference, a large body of literature is concerned with the decomposition of causal effects under the rubric of mediation analysis. However, many variations are spurious in nature, including different phenomena throughout the applied sciences. Despite the statistical power to estimate correlations and the identification power to decompose causal effects, there is still little understanding of the properties of spurious associations and how they can be decomposed in terms of the underlying causal mechanisms. In this manuscript, we develop formal tools for decomposing spurious variations in both Markovian and Semi-Markovian models. We prove the first results that allow a non-parametric decomposition of spurious effects and provide sufficient conditions for the identification of such decompositions. The described approach has several applications, ranging from explainable and fair AI to questions in epidemiology and medicine, and we empirically demonstrate its use.

Recent progress in scaling up large language models has shown impressive capabilities in performing few-shot learning across a wide range of natural language tasks. However, a key limitation is that these language models fundamentally lack grounding to visual perception - a crucial attribute needed to extend to real world tasks such as in visual-question answering and robotics. While prior works have largely connected image to text through pretraining or fine-tuning, learning such alignments are generally costly due to a combination of curating massive datasets and large computational burdens. In order to resolve these limitations, we propose a simple yet effective approach called Language-Quantized AutoEncoder (LQAE), a modification of VQ-VAE that learns to align text-image data in an unsupervised manner by leveraging pretrained language model denoisers (e.g., BERT). Our main idea is to encode images as sequences of text tokens by directly quantizing image embeddings using a pretrained language codebook. We then feed a masked version of the quantized embeddings into a BERT to reconstruct the original input. By doing so, LQAE learns to represent similar images with similar clusters of text tokens, thereby aligning these two modalities without the use of aligned text-image pairs. We show LQAE learns text-aligned image tokens that enable few-shot multi-modal learning with large language models, outperforming baseline methods in tasks such as image classification and VQA while requiring as few as 1-10 image-text pairs.

We present a rigorous methodology for auditing differentially private machine learning by adding multiple carefully designed examples called canaries. We take a first principles approach based on three key components. First, we introduce Lifted Differential Privacy (LiDP) that expands the definition of differential privacy to handle randomized datasets. This gives us the freedom to design randomized canaries. Second, we audit LiDP by trying to distinguish between the model trained with $K$ canaries versus $K-1$ canaries in the dataset, leaving one canary out. By drawing the canaries i.i.d., LiDP can leverage the symmetry in the design and reuse each privately trained model to run multiple statistical tests, one for each canary. Third, we introduce novel confidence intervals that take advantage of the multiple test statistics by adapting to the empirical higher-order correlations. Together, this new recipe demonstrates significant improvements in sample complexity, both theoretically and empirically, using synthetic and real data. Further, recent advances in designing stronger canaries can be readily incorporated in the new framework.

We study the problem of robustly estimating the mean or location parameter without moment assumptions. Known computationally efficient algorithms rely on strong distributional assumptions, such as sub-Gaussianity, or (certifiably) bounded moments. Moreover, the guarantees that they achieve in the heavy-tailed setting are weaker than those for sub-Gaussian distributions with known covariance. In this work, we show that such a tradeoff, between error guarantees and heavy-tails, is not necessary for symmetric distributions. We show that for a large class of symmetric distributions, the same error as in the Gaussian setting can be achieved efficiently. The distributions we study include products of arbitrary symmetric one-dimensional distributions, such as product Cauchy distributions, as well as elliptical distributions, a vast generalization of the Gaussian distribution. For product distributions and elliptical distributions with known scatter (covariance) matrix, we show that given an $\varepsilon$-corrupted sample, we can with probability at least $1-\delta$ estimate its location up to error $O(\varepsilon \sqrt{\log(1/\varepsilon)})$ using $\tfrac{d\log(d) + \log(1/\delta)}{\varepsilon^2 \log(1/\varepsilon)}$ samples. This result matches the best-known guarantees for the Gaussian distribution and known SQ lower bounds (up to the $\log(d)$ factor). For elliptical distributions with unknown scatter (covariance) matrix, we propose a sequence of efficient algorithms that approaches this optimal error. Specifically, for every $k \in \mathbb{N}$, we design an estimator using time and samples $\tilde{O}({d^k})$ achieving error $O(\varepsilon^{1-\frac{1}{2k}})$. This matches the error and running time guarantees when assuming certifiably bounded moments of order up to $k$. For unknown covariance, such error bounds of $o(\sqrt{\varepsilon})$ are not even known for (general) sub-Gaussian distributions. Our algorithms are based on a generalization of the well-known filtering technique [DK22]. More specifically, we show how this machinery can be combined with Huber-loss-based techniques to work with projections of the noise that behave more nicely than the initial noise. Moreover, we show how sum-of-squares proofs can be used to obtain algorithmic guarantees even for distributions without a first moment. We believe that this approach may find other applications in future works.

Labeling neural network submodules with human-legible descriptions is useful for many downstream tasks: such descriptions can surface failures, guide interventions, and perhaps even explain important model behaviors. To date, most mechanistic descriptions of trained networks have involved small models, narrowly delimited phenomena, and large amounts of human labor. Labeling all human-interpretable sub-computations in models of increasing size and complexity will almost certainly require tools that can generate and validate descriptions automatically. Recently, techniques that use learned models in-the-loop for labeling have begun to gain traction, but methods for evaluating their efficacy are limited and ad-hoc. How should we validate and compare open-ended labeling tools? This paper introduces FIND (Function INterpretation and Description), a benchmark suite for evaluating the building blocks of automated interpretability methods. FIND contains functions that resemble components of trained neural networks, and accompanying descriptions of the kind we seek to generate. The functions are procedurally constructed across textual and numeric domains, and involve a range of real-world complexities, including noise, composition, approximation, and bias. We evaluate methods that use pretrained language models (LMs) to produce code-based and natural language descriptions of function behavior. Additionally, we introduce a new interactive method in which an Automated Interpretability Agent (AIA) generates function descriptions. We find that an AIA, built with an off-the-shelf LM augmented with black-box access to functions, can sometimes infer function structure—acting as a scientist by forming hypotheses, proposing experiments, and updating descriptions in light of new data. However, FIND also reveals that LM-based descriptions capture global function behavior while missing local details. These results suggest that FIND will be useful for characterizing the performance of more sophisticated interpretability methods before they are applied to real-world models.

We propose a family of recursive cutting-plane algorithms to solve feasibility problems with constrained memory, which can also be used for first-order convex optimization. Precisely, in order to find a point within a ball of radius $\epsilon$ with a separation oracle in dimension $d$---or to minimize $1$-Lipschitz convex functions to accuracy $\epsilon$ over the unit ball---our algorithms use $\mathcal O(\frac{d^2}{p}\ln \frac{1}{\epsilon})$ bits of memory, and make $\mathcal O((C\frac{d}{p}\ln \frac{1}{\epsilon})^p)$ oracle calls. The family is parametrized by $p\in[d]$ and provides an oracle-complexity/memory trade-off in the sub-polynomial regime $\ln\frac{1}{\epsilon}\gg\ln d$. While several works gave lower-bound trade-offs (impossibility results)---we explicit here their dependence with $\ln\frac{1}{\epsilon}$, showing that these also hold in any sub-polynomial regime---to the best of our knowledge this is the first class of algorithms that provides a positive trade-off between gradient descent and cutting-plane methods in any regime with $\epsilon\leq 1/\sqrt d$. The algorithms divide the $d$ variables into $p$ blocks and optimize over blocks sequentially, with approximate separation vectors constructed using a variant of Vaidya's method. In the regime $\epsilon \leq d^{-\Omega(d)}$, our algorithm with $p=d$ achieves the information-theoretic optimal memory usage and improves the oracle-complexity of gradient descent.

What are the functionals of the reward that can be computed and optimized exactly in Markov Decision Processes? In the finite-horizon, undiscounted setting, Dynamic Programming (DP) can only handle these operations efficiently for certain classes of statistics. We summarize the characterization of these classes for policy evaluation, and give a new answer for the planning problem. Interestingly, we prove that only generalized means can be optimized exactly, even in the more general framework of Distributional Reinforcement Learning (DistRL). DistRL permits, however, to evaluate other functionals approximately. We provide error bounds on the resulting estimators, and discuss the potential of this approach as well as its limitations. These results contribute to advancing the theory of Markov Decision Processes by examining overall characteristics of the return, and particularly risk-conscious strategies.

By encoding time series as a string of numerical digits, we can frame time series forecasting as next-token prediction in text. Developing this approach, we find that large language models (LLMs) such as GPT-3 and LLaMA-2 can surprisingly zero-shot extrapolate time series at a level comparable to or exceeding the performance of purpose-built time series models trained on the downstream tasks. To facilitate this performance, we propose procedures for effectively tokenizing time series data and converting discrete distributions over tokens into highly flexible densities over continuous values. We argue the success of LLMs for time series stems from their ability to naturally represent multimodal distributions, in conjunction with biases for simplicity, and repetition, which align with the salient features in many time series, such as repeated seasonal trends. We also show how LLMs can naturally handle missing data without imputation through non-numerical text, accommodate textual side information, and answer questions to help explain predictions. While we find that increasing model size generally improves performance on time series, we show GPT-4 can perform worse than GPT-3 because of how it tokenizes numbers, and poor uncertainty calibration, which is likely the result of alignment interventions such as RLHF.

Different outdoor illumination conditions drastically alter the appearance of urban scenes, and they can harm the performance of image-based robot perception systems if not seen during training. Camera simulation provides a cost-effective solution to create a large dataset of images captured under different lighting conditions. Towards this goal, we propose LightSim, a neural lighting camera simulation system that enables diverse, realistic, and controllable data generation. LightSim automatically builds lighting-aware digital twins at scale from collected raw sensor data and decomposes the scene into dynamic actors and static background with accurate geometry, appearance, and estimated scene lighting. These digital twins enable actor insertion, modification, removal, and rendering from a new viewpoint, all in a lighting-aware manner. LightSim then combines physically-based and learnable deferred rendering to perform realistic relighting of modified scenes, such as altering the sun location and modifying the shadows or changing the sun brightness, producing spatially- and temporally-consistent camera videos. Our experiments show that LightSim generates more realistic relighting results than prior work. Importantly, training perception models on data generated by LightSim can significantly improve their performance. Our project page is available at https://waabi.ai/lightsim/.

Spiking Neural Networks (SNNs) have recently become more popular as a biologically plausible substitute for traditional Artificial Neural Networks (ANNs). SNNs are cost-efficient and deployment-friendly because they process input in both spatial and temporal manner using binary spikes. However, we observe that the information capacity in SNNs is affected by the number of timesteps, leading to an accuracy-efficiency tradeoff. In this work, we study a fine-grained adjustment of the number of timesteps in SNNs. Specifically, we treat the number of timesteps as a variable conditioned on different input samples to reduce redundant timesteps for certain data. We call our method Spiking Early-Exit Neural Networks (**SEENNs**). To determine the appropriate number of timesteps, we propose SEENN-I which uses a confidence score thresholding to filter out the uncertain predictions, and SEENN-II which determines the number of timesteps by reinforcement learning. Moreover, we demonstrate that SEENN is compatible with both the directly trained SNN and the ANN-SNN conversion. By dynamically adjusting the number of timesteps, our SEENN achieves a remarkable reduction in the average number of timesteps during inference. For example, our SEENN-II ResNet-19 can achieve **96.1**\% accuracy with an average of **1.08** timesteps on the CIFAR-10 test dataset. Code is shared at https://github.com/Intelligent-Computing-Lab-Yale/SEENN.

This paper theoretically explains the intuition that simple concepts are more likely to be learned by deep neural networks (DNNs) than complex concepts. In fact, recent studies have observed [24, 15] and proved [26] the emergence of interactive concepts in a DNN, i.e., it is proven that a DNN usually only encodes a small number of interactive concepts, and can be considered to use their interaction effects to compute inference scores. Each interactive concept is encoded by the DNN to represent the collaboration between a set of input variables. Therefore, in this study, we aim to theoretically explain that interactive concepts involving more input variables (i.e., more complex concepts) are more difficult to learn. Our finding clarifies the exact conceptual complexity that boosts the learning difficulty.

Standard conformal prediction methods provide a marginal coverage guarantee, which means that for a random test point, the conformal prediction set contains the true label with a user-specified probability. In many classification problems, we would like to obtain a stronger guarantee--that for test points of a specific class, the prediction set contains the true label with the same user-chosen probability. For the latter goal, existing conformal prediction methods do not work well when there is a limited amount of labeled data per class, as is often the case in real applications where the number of classes is large. We propose a method called clustered conformal prediction that clusters together classes having "similar" conformal scores and performs conformal prediction at the cluster level. Based on empirical evaluation across four image data sets with many (up to 1000) classes, we find that clustered conformal typically outperforms existing methods in terms of class-conditional coverage and set size metrics.

Offline Reinforcement Learning (RL) aims to address the challenge of distribution shift between the dataset and the learned policy, where the value of out-of-distribution (OOD) data may be erroneously estimated due to overgeneralization. It has been observed that a considerable portion of the benefits derived from the conservative terms designed by existing offline RL approaches originates from their impact on the learned representation. This observation prompts us to scrutinize the learning dynamics of offline RL, formalize the process of generalization, and delve into the prevalent overgeneralization issue in offline RL. We then investigate the potential to rein the generalization from the representation perspective to enhance offline RL. Finally, we present Representation Distinction (RD), an innovative plug-in method for improving offline RL algorithm performance by explicitly differentiating between the representations of in-sample and OOD state-action pairs generated by the learning policy. Considering scenarios in which the learning policy mirrors the behavioral policy and similar samples may be erroneously distinguished, we suggest a dynamic adjustment mechanism for RD based on an OOD data generator to prevent data representation collapse and further enhance policy performance. We demonstrate the efficacy of our approach by applying RD to specially-designed backbone algorithms and widely-used offline RL algorithms. The proposed RD method significantly improves their performance across various continuous control tasks on D4RL datasets, surpassing several state-of-the-art offline RL algorithms.

Meta-learning enables quick adaptation of machine learning models to new tasks with limited data. While tasks could come from varying distributions in reality, most of the existing meta-learning methods consider both training and testing tasks as from the same uni-component distribution, overlooking two critical needs of a practical solution: (1) the various sources of tasks may compose a multi-component mixture distribution, and (2) novel tasks may come from a distribution that is unseen during meta-training. In this paper, we demonstrate these two challenges can be solved jointly by modeling the density of task instances. We develop a meta-training framework underlain by a novel Hierarchical Gaussian Mixture based Task Generative Model (HTGM). HTGM extends the widely used empirical process of sampling tasks to a theoretical model, which learns task embeddings, fits the mixture distribution of tasks, and enables density-based scoring of novel tasks. The framework is agnostic to the encoder and scales well with large backbone networks. The model parameters are learned end-to-end by maximum likelihood estimation via an Expectation-Maximization (EM) algorithm. Extensive experiments on benchmark datasets indicate the effectiveness of our method for both sample classification and novel task detection.

Soft prompt tuning achieves superior performances across a wide range of few-shot tasks. However, the performances of prompt tuning can be highly sensitive to the initialization of the prompts. We have also empirically observed that conventional prompt tuning methods cannot encode and learn sufficient task-relevant information from prompt tokens. In this work, we develop an information-theoretic framework that formulates soft prompt tuning as maximizing the mutual information between prompts and other model parameters (or encoded representations). This novel view helps us to develop a more efficient, accurate and robust soft prompt tuning method, InfoPrompt. With this framework, we develop two novel mutual information based loss functions, to (i) explore proper prompt initialization for the downstream tasks and learn sufficient task-relevant information from prompt tokens and (ii) encourage the output representation from the pretrained language model to be more aware of the task-relevant information captured in the learnt prompts. Extensive experiments validate that InfoPrompt can significantly accelerate the convergence of the prompt tuning and outperform traditional prompt tuning methods. Finally, we provide a formal theoretical result to show that a gradient descent type algorithm can be used to train our mutual information loss.

This paper introduces the novel concept of few-shot weakly supervised learning for pathology Whole Slide Image (WSI) classification, denoted as FSWC. A solution is proposed based on prompt learning and the utilization of a large language model, GPT-4. Since a WSI is too large and needs to be divided into patches for processing, WSI classification is commonly approached as a Multiple Instance Learning (MIL) problem. In this context, each WSI is considered a bag, and the obtained patches are treated as instances. The objective of FSWC is to classify both bags and instances with only a limited number of labeled bags. Unlike conventional few-shot learning problems, FSWC poses additional challenges due to its weak bag labels within the MIL framework. Drawing inspiration from the recent achievements of vision-language models (V-L models) in downstream few-shot classification tasks, we propose a two-level prompt learning MIL framework tailored for pathology, incorporating language prior knowledge. Specifically, we leverage CLIP to extract instance features for each patch, and introduce a prompt-guided pooling strategy to aggregate these instance features into a bag feature. Subsequently, we employ a small number of labeled bags to facilitate few-shot prompt learning based on the bag features. Our approach incorporates the utilization of GPT-4 in a question-and-answer mode to obtain language prior knowledge at both the instance and bag levels, which are then integrated into the instance and bag level language prompts. Additionally, a learnable component of the language prompts is trained using the available few-shot labeled data. We conduct extensive experiments on three real WSI datasets encompassing breast cancer, lung cancer, and cervical cancer, demonstrating the notable performance of the proposed method in bag and instance classification. All codes will be made publicly accessible.

Contrastive learning (CL) has been widely investigated with various learning mechanisms and achieves strong capability in learning representations of data in a self-supervised manner using unlabeled data. A common fashion of contrastive learning on this line is employing mega-sized encoders to achieve comparable performance as the supervised learning counterpart. Despite the success of the labelless training, current contrastive learning algorithms *failed* to achieve good performance with lightweight (compact) models, e.g., MobileNet, while the requirements of the heavy encoders impede the energy-efficient computation, especially for resource-constrained AI applications. Motivated by this, we propose a new self-supervised CL scheme, named SACL-XD, consisting of two technical components, **S**limmed **A**symmetrical **C**ontrastive **L**earning (SACL) and **Cross**-**D**istillation (XD), which collectively enable efficient CL with compact models. While relevant prior works employed a strong pre-trained model as the teacher of unsupervised knowledge distillation to a lightweight encoder, our proposed method trains CL models from scratch and outperforms them even without such an expensive requirement. Compared to the SoTA lightweight CL training (distillation) algorithms, SACL-XD achieves 1.79% ImageNet-1K accuracy improvement on MobileNet-V3 with 64$\times$ training FLOPs reduction.

A machine learning model is traditionally considered robust if its prediction remains (almost) constant under input perturbations with small norm. However, real-world tasks like molecular property prediction or point cloud segmentation have inherent equivariances, such as rotation or permutation equivariance. In such tasks, even perturbations with large norm do not necessarily change an input's semantic content. Furthermore, there are perturbations for which a model's prediction explicitly needs to change. For the first time, we propose a sound notion of adversarial robustness that accounts for task equivariance. We then demonstrate that provable robustness can be achieved by (1) choosing a model that matches the task's equivariances (2) certifying traditional adversarial robustness. Certification methods are, however, unavailable for many models, such as those with continuous equivariances. We close this gap by developing the framework of equivariance-preserving randomized smoothing, which enables architecture-agnostic certification. We additionally derive the first architecture-specific graph edit distance certificates, i.e. sound robustness guarantees for isomorphism equivariant tasks like node classification. Overall, a sound notion of robustness is an important prerequisite for future work at the intersection of robust and geometric machine learning.

We give the first polynomial-time algorithm for the testable learning of halfspaces in the presence of adversarial label noise under the Gaussian distribution. In the recently introduced testable learning model, one is required to produce a tester-learner such that if the data passes the tester, then one can trust the output of the robust learner on the data. Our tester-learner runs in time $\text{poly}(d/\epsilon)$ and outputs a halfspace with misclassification error $O(\text{opt})+\epsilon$, where $\text{opt}$ is the 0-1 error of the best fitting halfspace. At a technical level, our algorithm employs an iterative soft localization technique enhanced with appropriate testers to ensure that the data distribution is sufficiently similar to a Gaussian. Finally, our algorithm can be readily adapted to yield an efficient and testable active learner requiring only $d ~ \text{polylog}(1/\epsilon)$ labeled examples.