NeurIPS 2023 Papers

Data and code working together is fundamental to machine learning (ML), but the context around datasets and interactions between datasets and code are in general captured only rudimentarily. Context such as how the dataset was prepared and created, what source data were used, what code was used in processing, how the dataset evolved, and where it has been used and reused can provide much insight, but this information is often poorly documented. That is unfortunate since it makes datasets into black-boxes with potentially hidden characteristics that have downstream consequences. We argue that making dataset preparation more accessible and dataset usage easier to record and document would have significant benefits for the ML community: it would allow for greater diversity in datasets by inviting modification to published sources, simplify use of alternative datasets and, in doing so, make results more transparent and robust, while allowing for all contributions to be adequately credited. We present a platform, Renku, designed to support and encourage such sustainable development and use of data, datasets, and code, and we demonstrate its benefits through a few illustrative projects which span the spectrum from dataset creation to dataset consumption and showcasing.

Input-Convex Neural Networks (ICNNs) are networks that guarantee convexity in their input-output mapping. These networks have been successfully applied for energy-based modelling, optimal transport problems and learning invariances. The convexity of ICNNs is achieved by using non-decreasing convex activation functions and non-negative weights. Because of these peculiarities, previous initialisation strategies, which implicitly assume centred weights, are not effective for ICNNs. By studying signal propagation through layers with non-negative weights, we are able to derive a principled weight initialisation for ICNNs. Concretely, we generalise signal propagation theory by removing the assumption that weights are sampled from a centred distribution. In a set of experiments, we demonstrate that our principled initialisation effectively accelerates learning in ICNNs and leads to better generalisation. Moreover, we find that, in contrast to common belief, ICNNs can be trained without skip-connections when initialised correctly. Finally, we apply ICNNs to a real-world drug discovery task and show that they allow for more effective molecular latent space exploration.

Modeling complex dynamical systems poses significant challenges, with traditional methods struggling to work across a variety of systems and scale to high-dimensional dynamics. In response, we present DynaDojo, a novel benchmarking platform designed for data-driven dynamical system identification. DynaDojo enables comprehensive evaluation of how an algorithm's performance scales across three key dimensions: (1) the number of training samples provided, (2) the complexity of the dynamical system being modeled, and (3) the training samples required to achieve a target error threshold. Furthermore, DynaDojo enables studying out-of-distribution generalization (by providing multiple test conditions for each system) and active learning (by supporting closed-loop control). Through its user-friendly and easily extensible API, DynaDojo accommodates a wide range of user-defined $\texttt{Algorithms}$, $\texttt{Systems}$, and $\texttt{Challenges}$ (scaling metrics). The platform also prioritizes resource-efficient training for running on a cluster. To showcase its utility, in DynaDojo $\texttt{0.9}$, we include implementations of 7 baseline algorithms and 20 dynamical systems, along with many demo notebooks. This work aspires to make DynaDojo a unifying benchmarking platform for system identification, paralleling the role of OpenAI’s Gym in reinforcement learning.

Neural Radiance Fields (NeRFs) have achieved impressive results in novel view synthesis and surface reconstruction tasks. However, their performance suffers under challenging scenarios with sparse input views. We present CorresNeRF, a novel method that leverages image correspondence priors computed by off-the-shelf methods to supervise NeRF training. We design adaptive processes for augmentation and filtering to generate dense and high-quality correspondences. The correspondences are then used to regularize NeRF training via the correspondence pixel reprojection and depth loss terms. We evaluate our methods on novel view synthesis and surface reconstruction tasks with density-based and SDF-based NeRF models on different datasets. Our method outperforms previous methods in both photometric and geometric metrics. We show that this simple yet effective technique of using correspondence priors can be applied as a plug-and-play module across different NeRF variants. The project page is at https://yxlao.github.io/corres-nerf/.

Temporal networks are widely used as abstract graph representations for real-world dynamic systems. Indeed, recognizing the network evolution states is crucial in understanding and analyzing temporal networks. For instance, social networks will generate the clustering and formation of tightly-knit groups or communities over time, relying on the triadic closure theory. However, the existing methods often struggle to account for the time-varying nature of these network structures, hindering their performance when applied to networks with complex evolution states. To mitigate this problem, we propose a novel framework called ESSEN, an Evolution StateS awarE Network, to measure temporal network evolution using von Neumann entropy and thermodynamic temperature. The developed framework utilizes a von Neumann entropy aware attention mechanism and network evolution state contrastive learning in the graph encoding. In addition, it employs a unique decoder the so-called Mixture of Thermodynamic Experts (MoTE) for decoding. ESSEN extracts local and global network evolution information using thermodynamic features and adaptively recognizes the network evolution states. Moreover, the proposed method is evaluated on link prediction tasks under both transductive and inductive settings, with the corresponding results demonstrating its effectiveness compared to various state-of-the-art baselines.

Given varying prompts regarding a factoid question, can a large language model (LLM) reliably generate factually correct answers? Existing LLMs may generate distinct responses for different prompts. In this paper, we study the problem of quantifying knowledge contained in an LLM regarding a given set of facts. We propose KaRR, a statistical approach to assess factual knowledge for LLMs. The main idea is to estimate the ratio of LLM generating text corresponding to the answer entity given diverse prompts of the subject and the querying relation, versus it generating by random chances. Our assessment suite contains a comprehensive set of 994,123 entities and 600 relations, with 1,395,905 text aliases. We use our method to evaluate 20 LLMs of various sizes, including LLaMA, Alpaca, OPT, etc. Experiments show that our results have a strong correlation (0.43 Kendall's $\tau$) with the results of human assessment on LLMs. Our results reveal that the knowledge in LLMs with the same backbone architecture adheres to the scaling law, while tuning on instruction-following data sometimes compromises the model's capability to generate factually correct text reliably.

Despite a great deal of research, it is still not well-understood why trained neural networks are highly vulnerable to adversarial examples. In this work we focus on two-layer neural networks trained using data which lie on a low dimensional linear subspace. We show that standard gradient methods lead to non-robust neural networks, namely, networks which have large gradients in directions orthogonal to the data subspace, and are susceptible to small adversarial $L_2$-perturbations in these directions. Moreover, we show that decreasing the initialization scale of the training algorithm, or adding $L_2$ regularization, can make the trained network more robust to adversarial perturbations orthogonal to the data.

Sampling in discrete spaces, with critical applications in simulation and optimization, has recently been boosted by significant advances in gradient-based approaches that exploit modern accelerators like GPUs. However, two key challenges are hindering further advancement in research on discrete sampling. First, since there is no consensus on experimental settings and evaluation setups, the empirical results in different research papers are often not comparable. Second, implementing samplers and target distributions often requires a nontrivial amount of effort in terms of calibration and parallelism. To tackle these challenges, we propose DISCS (DISCrete Sampling), a tailored package and benchmark that supports unified and efficient experiment implementation and evaluations for discrete sampling in three types of tasks: sampling from classical graphical models and energy based generative models, and sampling for solving combinatorial optimization. Throughout the comprehensive evaluations in DISCS, we gained new insights into scalability, design principles for proposal distributions, and lessons for adaptive sampling design. DISCS efficiently implements representative discrete samplers in existing research works as baselines and offers a simple interface that researchers can conveniently add new discrete samplers and directly compare their performance with the benchmark result in a calibrated setup.

We study the problem of learning mixtures of linear classifiers under Gaussian covariates. Given sample access to a mixture of $r$ distributions on $\mathbb{R}^n$ of the form $(\mathbf{x},y_{\ell})$, $\ell \in [r]$, where $\mathbf{x}\sim\mathcal{N}(0,\mathbf{I}_n)$ and $y_\ell=\mathrm{sign}(\langle\mathbf{v}_{\ell},\mathbf{x}\rangle)$ for an unknown unit vector $\mathbf{v}_{\ell}$, the goal is to learn the underlying distribution in total variation distance. Our main result is a Statistical Query (SQ) lower bound suggesting that known algorithms for this problem are essentially best possible, even for the special case of uniform mixtures. In particular, we show that the complexity of any SQ algorithm for the problem is $n^{\mathrm{poly}(1/\Delta) \log(r)}$, where $\Delta$ is a lower bound on the pairwise $\ell_2$-separation between the $\mathbf{v}_{\ell}$'s. The key technical ingredient underlying our result is a new construction of spherical designs on the unit sphere that may be of independent interest.

In this paper, we propose to approximate the softmax output, which is the key product of the attention mechanism, to reduce its activation memory usage when training attention-based networks (aka Transformers). During the forward pass of the network, the proposed softmax output approximation method stores only a small fraction of the entire softmax output required for back-propagation and evicts the rest of the softmax output from memory. Then, during the backward pass, the evicted softmax activation output is approximated to compose the gradient to perform back-propagation for model training. Considering most attention-based models heavily rely on the softmax-based attention module that usually takes one of the biggest portions of the network, approximating the softmax activation output can be a simple yet effective way to decrease the training memory requirement of many attention-based networks. The experiment with various attention-based models and relevant tasks, i.e., machine translation, text classification, and sentiment analysis, shows that it curtails the activation memory usage of the softmax-based attention module by up to 84% (6.2× less memory) in model training while achieving comparable or better performance, e.g., up to 5.4% higher classification accuracy.

Sequence prediction on temporal data requires the ability to understand compositional structures of multi-level semantics beyond individual and contextual properties of parts. The task of temporal action segmentation remains challenging for the reason, aiming at translating an untrimmed activity video into a sequence of action segments. This paper addresses the problem by introducing an effective activity grammar to guide neural predictions for temporal action segmentation. We propose a novel grammar induction algorithm, dubbed KARI, that extracts a powerful context-free grammar from action sequence data. We also develop an efficient generalized parser, dubbed BEP, that transforms frame-level probability distributions into a reliable sequence of actions according to the induced grammar with recursive rules. Our approach can be combined with any neural network for temporal action segmentation to enhance the sequence prediction and discover its compositional structure. Experimental results demonstrate that our method significantly improves temporal action segmentation in terms of both performance and interpretability on two standard benchmarks, Breakfast and 50 Salads.

Binary Balanced Tree Recursive Neural Networks (BBT-RvNNs) enforce sequence composition according to a preset balanced binary tree structure. Thus, their non-linear recursion depth (which is the tree depth) is just $\log_2 n$ ($n$ being the sequence length). Such logarithmic scaling makes BBT-RvNNs efficient and scalable on long sequence tasks such as Long Range Arena (LRA). However, such computational efficiency comes at a cost because BBT-RvNNs cannot solve simple arithmetic tasks like ListOps. On the flip side, RvNN models (e.g., Beam Tree RvNN) that do succeed on ListOps (and other structure-sensitive tasks like formal logical inference) are generally several times more expensive (in time and space) than even Recurrent Neural Networks. In this paper, we introduce a novel framework --- Recursion in Recursion (RIR) to strike a balance between the two sides - getting some of the benefits from both worlds. In RIR, we use a form of two-level nested recursion - where the outer recursion is a $k$-ary balanced tree model with another recursive model (inner recursion) implementing its cell function. For the inner recursion, we choose Beam Tree RvNNs. To adjust Beam Tree RvNNs within RIR we also propose a novel strategy of beam alignment. Overall, this entails that the total recursive depth in RIR is upper-bounded by $k \log_k n$. Our best RIR-based model is the first model that demonstrates high ($\geq 90\%$) length-generalization performance on ListOps while at the same time being scalable enough to be trainable on long sequence inputs from LRA (it can reduce the memory usage of the original Beam Tree RvNN by hundreds of times). Moreover, in terms of accuracy in the LRA language tasks, it performs competitively with Structured State Space Models (SSMs) without any special initialization - outperforming Transformers by a large margin. On the other hand, while SSMs can marginally outperform RIR on LRA, they (SSMs) fail to length-generalize on ListOps. Our code is available at: https://github.com/JRC1995/BeamRecursionFamily/

Facial expression data is characterized by a significant imbalance, with most collected data showing happy or neutral expressions and fewer instances of fear or disgust. This imbalance poses challenges to facial expression recognition (FER) models, hindering their ability to fully understand various human emotional states. Existing FER methods typically report overall accuracy on highly imbalanced test sets but exhibit low performance in terms of the mean accuracy across all expression classes. In this paper, our aim is to address the imbalanced FER problem. Existing methods primarily focus on learning knowledge of minor classes solely from minor-class samples. However, we propose a novel approach to extract extra knowledge related to the minor classes from both major and minor class samples. Our motivation stems from the belief that FER resembles a distribution learning task, wherein a sample may contain information about multiple classes. For instance, a sample from the major class surprise might also contain useful features of the minor class fear. Inspired by that, we propose a novel method that leverages re-balanced attention maps to regularize the model, enabling it to extract transformation invariant information about the minor classes from all training samples. Additionally, we introduce re-balanced smooth labels to regulate the cross-entropy loss, guiding the model to pay more attention to the minor classes by utilizing the extra information regarding the label distribution of the imbalanced training data. Extensive experiments on different datasets and backbones show that the two proposed modules work together to regularize the model and achieve state-of-the-art performance under the imbalanced FER task. Code is available at https://github.com/zyh-uaiaaaa.

In recent years, significant attention in deep learning theory has been devoted to analyzing when models that interpolate their training data can still generalize well to unseen examples. Many insights have been gained from studying models with multiple layers of Gaussian random features, for which one can compute precise generalization asymptotics. However, few works have considered the effect of weight anisotropy; most assume that the random features are generated using independent and identically distributed Gaussian weights, and allow only for structure in the input data. Here, we use the replica trick from statistical physics to derive learning curves for models with many layers of structured Gaussian features. We show that allowing correlations between the rows of the first layer of features can aid generalization, while structure in later layers is generally detrimental. Our results shed light on how weight structure affects generalization in a simple class of solvable models.

Federated learning (FL) is usually performed on resource-constrained edge devices, e.g., with limited memory for the computation. If the required memory to train a model exceeds this limit, the device will be excluded from the training. This can lead to a lower accuracy as valuable data and computation resources are excluded from training, also causing bias and unfairness. The FL training process should be adjusted to such constraints. The state-of-the-art techniques propose training subsets of the FL model at constrained devices, reducing their resource requirements for training. However, these techniques largely limit the co-adaptation among parameters of the model and are highly inefficient, as we show: it is actually better to train a smaller (less accurate) model by the system where all the devices can train the model end-to-end than applying such techniques. We propose a new method that enables successive freezing and training of the parameters of the FL model at devices, reducing the training’s resource requirements at the devices while still allowing enough co-adaptation between parameters. We show through extensive experimental evaluation that our technique greatly improves the accuracy of the trained model (by 52.4 p.p. ) compared with the state of the art, efficiently aggregating the computation capacity available on distributed devices.

Trajectory prediction plays a vital role in the performance of autonomous driving systems, and prediction accuracy, such as average displacement error (ADE) or final displacement error (FDE), is widely used as a performance metric. However, a significant disparity exists between the accuracy of predictors on fixed datasets and driving performance when the predictors are used downstream for vehicle control, because of a dynamics gap. In the real world, the prediction algorithm influences the behavior of the ego vehicle, which, in turn, influences the behaviors of other vehicles nearby. This interaction results in predictor-specific dynamics that directly impacts prediction results. In fixed datasets, since other vehicles' responses are predetermined, this interaction effect is lost, leading to a significant dynamics gap. This paper studies the overlooked significance of this dynamics gap. We also examine several other factors contributing to the disparity between prediction performance and driving performance. The findings highlight the trade-off between the predictor's computational efficiency and prediction accuracy in determining real-world driving performance. In summary, an interactive, task-driven evaluation protocol for trajectory prediction is crucial to capture its effectiveness for autonomous driving. Source code along with experimental settings is available online (https://whatmatters23.github.io/).

Graph Neural Networks (GNNs) have been shown to achieve remarkable performance on node classification tasks by exploiting both graph structures and node features. The majority of existing GNNs rely on the implicit homophily assumption. Recent studies have demonstrated that GNNs may struggle to model heterophilous graphs where nodes with different labels are more likely connected. To address this issue, we propose a generic GNN applicable to both homophilous and heterophilous graphs, namely Low-Rank Graph Neural Network (LRGNN). Our analysis demonstrates that a signed graph's global label relationship matrix has a low rank. This insight inspires us to predict the label relationship matrix by solving a robust low-rank matrix approximation problem, as prior research has proven that low-rank approximation could achieve perfect recovery under certain conditions. The experimental results reveal that the solution bears a strong resemblance to the label relationship matrix, presenting two advantages for graph modeling: a block diagonal structure and varying distributions of within-class and between-class entries.

Deep learning methods have gained popularity in high energy physics for fast modeling of particle showers in detectors. Detailed simulation frameworks such as the gold standard \textsc{Geant4} are computationally intensive, and current deep generative architectures work on discretized, lower resolution versions of the detailed simulation. The development of models that work at higher spatial resolutions is currently hindered by the complexity of the full simulation data, and by the lack of simpler, more interpretable benchmarks. Our contribution is \textsc{SUPA}, the SUrrogate PArticle propagation simulator, an algorithm and software package for generating data by simulating simplified particle propagation, scattering and shower development in matter. The generation is extremely fast and easy to use compared to \textsc{Geant4}, but still exhibits the key characteristics and challenges of the detailed simulation. The proposed simulator generates thousands of particle showers per second on a desktop machine, a speed up of up to 6 orders of magnitudes over \textsc{Geant4}, and stores detailed geometric information about the shower propagation. \textsc{\textsc{SUPA}} provides much greater flexibility for setting initial conditions and defining multiple benchmarks for the development of models. Moreover, interpreting particle showers as point clouds creates a connection to geometric machine learning and provides challenging and fundamentally new datasets for the field.

Deep Neural Networks and Reinforcement Learning methods have empirically shown great promise in tackling challenging combinatorial problems. In those methods a deep neural network is used as a solution generator which is then trained by gradient-based methods (e.g., policy gradient) to successively obtain better solution distributions. In this work we introduce a novel theoretical framework for analyzing the effectiveness of such methods. We ask whether there exist generative models that (i) are expressive enough to generate approximately optimal solutions; (ii) have a tractable, i.e, polynomial in the size of the input, number of parameters; (iii) their optimization landscape is benign in the sense that it does not contain sub-optimal stationary points. Our main contribution is a positive answer to this question. Our result holds for a broad class of combinatorial problems including Max- and Min-Cut, Max-$k$-CSP, Maximum-Weight-Bipartite-Matching, and the Traveling Salesman Problem. As a byproduct of our analysis we introduce a novel regularization process over vanilla gradient descent and provide theoretical and experimental evidence that it helps address vanishing-gradient issues and escape bad stationary points.