NeurIPS 2023 Papers

In this paper, we explore the structure of the penultimate Gram matrix in deep neural networks, which contains the pairwise inner products of outputs corresponding to a batch of inputs. In several architectures it has been observed that this Gram matrix becomes degenerate with depth at initialization, which dramatically slows training. Normalization layers, such as batch or layer normalization, play a pivotal role in preventing the rank collapse issue. Despite promising advances, the existing theoretical results do not extend to layer normalization, which is widely used in transformers, and can not quantitatively characterize the role of non-linear activations. To bridge this gap, we prove that layer normalization, in conjunction with activation layers, biases the Gram matrix of a multilayer perceptron towards the identity matrix at an exponential rate with depth at initialization. We quantify this rate using the Hermite expansion of the activation function.

The ocean is a crucial component of the Earth's system. It profoundly influences human activities and plays a critical role in climate regulation. Our understanding has significantly improved over the last decades with the advent of satellite remote sensing data, allowing us to capture essential sea surface quantities over the globe, e.g., sea surface height (SSH). Despite their ever-increasing abundance, ocean satellite data presents challenges for information extraction due to their sparsity and irregular sampling, signal complexity, and noise. Machine learning (ML) techniques have demonstrated their capabilities in dealing with large-scale, complex signals. Therefore we see an opportunity for these ML models to harness the full extent of the information contained in ocean satellite data. However, data representation and relevant evaluation metrics can be the defining factors when determining the success of applied ML. The processing steps from the raw observation data to a ML-ready state and from model outputs to interpretable quantities require domain expertise, which can be a significant barrier to entry for ML researchers. In addition, imposing fixed processing steps, like committing to specific variables, regions, and geometries, will narrow the scope of ML models and their potential impact on real-world applications. OceanBench is a unifying framework that provides standardized processing steps that comply with domain-expert standards. It is designed with a flexible and pedagogical abstraction: it a) provides plug-and-play data and pre-configured pipelines for ML researchers to benchmark their models w.r.t. ML and domain-related baselines and b) provides a transparent and configurable framework for researchers to customize and extend the pipeline for their tasks. In this work, we demonstrate the OceanBench framework through a first edition dedicated to SSH interpolation challenges. We provide datasets and ML-ready benchmarking pipelines for the long-standing problem of interpolating observations from simulated ocean satellite data, multi-modal and multi-sensor fusion issues, and transfer-learning to real ocean satellite observations. The OceanBench framework is available at https://github.com/jejjohnson/oceanbench and the dataset registry is available at https://github.com/quentinf00/oceanbench-data-registry.

Robots often rely on a repertoire of previously-learned motion policies for performing tasks of diverse complexities. When facing unseen task conditions or when new task requirements arise, robots must adapt their motion policies accordingly. In this context, policy optimization is the \emph{de facto} paradigm to adapt robot policies as a function of task-specific objectives. Most commonly-used motion policies carry particular structures that are often overlooked in policy optimization algorithms. We instead propose to leverage the structure of probabilistic policies by casting the policy optimization as an optimal transport problem. Specifically, we focus on robot motion policies that build on Gaussian mixture models (GMMs) and formulate the policy optimization as a Wassertein gradient flow over the GMMs space. This naturally allows us to constrain the policy updates via the $L^2$-Wasserstein distance between GMMs to enhance the stability of the policy optimization process. Furthermore, we leverage the geometry of the Bures-Wasserstein manifold to optimize the Gaussian distributions of the GMM policy via Riemannian optimization. We evaluate our approach on common robotic settings: Reaching motions, collision-avoidance behaviors, and multi-goal tasks. Our results show that our method outperforms common policy optimization baselines in terms of task success rate and low-variance solutions.

Pre-training is crucial in 3D-related fields such as autonomous driving where point cloud annotation is costly and challenging. Many recent studies on point cloud pre-training, however, have overlooked the issue of incompleteness, where only a fraction of the points are captured by LiDAR, leading to ambiguity during the training phase. On the other hand, images offer more comprehensive information and richer semantics that can bolster point cloud encoders in addressing the incompleteness issue inherent in point clouds. Yet, incorporating images into point cloud pre-training presents its own challenges due to occlusions, potentially causing misalignments between points and pixels. In this work, we propose PRED, a novel image-assisted pre-training framework for outdoor point clouds in an occlusion-aware manner. The main ingredient of our framework is a Birds-Eye-View (BEV) feature map conditioned semantic rendering, leveraging the semantics of images for supervision through neural rendering. We further enhance our model's performance by incorporating point-wise masking with a high mask ratio (95%). Extensive experiments demonstrate PRED's superiority over prior point cloud pre-training methods, providing significant improvements on various large-scale datasets for 3D perception tasks. Codes will be available at https://github.com/PRED4pc/PRED.

Neural Ordinary Differential Equations (NeuralODEs) present an attractive way to extract dynamical laws from time series data, as they bridge neural networks with the differential equation-based modeling paradigm of the physical sciences. However, these models often display long training times and suboptimal results, especially for longer duration data. While a common strategy in the literature imposes strong constraints to the NeuralODE architecture to inherently promote stable model dynamics, such methods are ill-suited for dynamics discovery as the unknown governing equation is not guaranteed to satisfy the assumed constraints. In this paper, we develop a new training method for NeuralODEs, based on synchronization and homotopy optimization, that does not require changes to the model architecture. We show that synchronizing the model dynamics and the training data tames the originally irregular loss landscape, which homotopy optimization can then leverage to enhance training. Through benchmark experiments, we demonstrate our method achieves competitive or better training loss while often requiring less than half the number of training epochs compared to other model-agnostic techniques. Furthermore, models trained with our method display better extrapolation capabilities, highlighting the effectiveness of our method.

The distillation of ranking models has become an important topic in both academia and industry. In recent years, several advanced methods have been proposed to tackle this problem, often leveraging ranking information from teacher rankers that is absent in traditional classification settings. To date, there is no well-established consensus on how to evaluate this class of models. Moreover, inconsistent benchmarking on a wide range of tasks and datasets make it difficult to assess or invigorate advances in this field. This paper first examines representative prior arts on ranking distillation, and raises three questions to be answered around methodology and reproducibility. To that end, we propose a systematic and unified benchmark, Ranking Distillation Suite (RD-Suite), which is a suite of tasks with 4 large real-world datasets, encompassing two major modalities (textual and numeric) and two applications (standard distillation and distillation transfer). RD-Suite consists of benchmark results that challenge some of the common wisdom in the field, and the release of datasets with teacher scores and evaluation scripts for future research. RD-Suite paves the way towards better understanding of ranking distillation, facilities more research in this direction, and presents new challenges.

Graph Neural Networks (GNNs) are a powerful class of machine learning models with applications in recommender systems, drug discovery, social network analysis, and computer vision. One challenge with their implementation is that GNNs often take large-scale graphs as inputs, which imposes significant computational/storage costs in the training and testing phases. In particular, the message passing operations of a GNN require multiplication of the graph adjacency matrix $A \in \mathbb{R}^{n \times n}$ and the data matrix $X \in \mathbb{R}^{n \times d}$, and the $O(n^2 d)$ time complexity can be prohibitive for large $n$. Thus, a natural question is whether it is possible to perform the GNN operations in (quasi-)linear time by avoiding the full computation of $A X$. To study this question, we consider the setting of a regression task on a two-layer Linear Graph Convolutional Network (GCN). We develop an efficient training algorithm based on (1) performing node subsampling, (2) estimating the leverage scores of $A X$ based on the subsampled graph, and (3) performing leverage score sampling on $A X$. We show that our proposed scheme learns the regression model observing only $O(nd\epsilon^{-2}\log n)$ entries of $A$ in time $O(nd^2 \epsilon^{-2}\log n)$, with the guarantee that the learned weights deviate by at most $\epsilon$ under the $\ell_2$ norm from the model learned using the entire adjacency matrix $A$. We present empirical results for regression problems on real-world graphs and show that our algorithm significantly outperforms other baseline sampling strategies that exploit the same number of observations.

Recently, the anchor acceleration, an acceleration mechanism distinct from Nesterov's, has been discovered for minimax optimization and fixed-point problems, but its mechanism is not understood well, much less so than Nesterov acceleration. In this work, we analyze continuous-time models of anchor acceleration. We provide tight, unified analyses for characterizing the convergence rate as a function of the anchor coefficient $\beta(t)$, thereby providing insight into the anchor acceleration mechanism and its accelerated $\mathcal{O}(1/k^2)$-convergence rate. Finally, we present an adaptive method inspired by the continuous-time analyses and establish its effectiveness through theoretical analyses and experiments.

Low-rank matrix factorization (LRMF) is a canonical problem in non-convex optimization, the objective function to be minimized is non-convex and even non-smooth, which makes the global convergence guarantee of gradient-based algorithm quite challenging. Recent work made a breakthrough on proving that standard gradient descent converges to the $\varepsilon$-global minima after $O( \frac{d \kappa^2}{\tau^2} {\rm ln} \frac{d \sigma_d}{\tau} + \frac{d \kappa^2}{\tau^2} {\rm ln} \frac{\sigma_d}{\varepsilon})$ iterations from small initialization with a very small learning rate (both are related to the small constant $\tau$). While the dependence of the convergence on the \textit{condition number} $\kappa$ and \textit{small learning rate} makes it not practical especially for ill-conditioned LRMF problem. In this paper, we show that precondition helps in accelerating the convergence and prove that the scaled gradient descent (ScaledGD) and its variant, alternating scaled gradient descent (AltScaledGD) converge to an $\varepsilon$-global minima after $O( {\rm ln} \frac{d}{\delta} + {\rm ln} \frac{d}{\varepsilon})$ iterations from general random initialization. Meanwhile, for small initialization as in gradient descent, both ScaledGD and AltScaledGD converge to $\varepsilon$-global minima after only $O({\rm ln} \frac{d}{\varepsilon})$ iterations. Furthermore, we prove that as a proximity to the alternating minimization, AltScaledGD converges faster than ScaledGD, its global convergence does not rely on small learning rate and small initialization, which certificates the advantages of AltScaledGD in LRMF.

Estimating 3D articulated shapes like animal bodies from monocular images is inherently challenging due to the ambiguities of camera viewpoint, pose, texture, lighting, etc. We propose ARTIC3D, a self-supervised framework to reconstruct per-instance 3D shapes from a sparse image collection in-the-wild. Specifically, ARTIC3D is built upon a skeleton-based surface representation and is further guided by 2D diffusion priors from Stable Diffusion. First, we enhance the input images with occlusions/truncation via 2D diffusion to obtain cleaner mask estimates and semantic features. Second, we perform diffusion-guided 3D optimization to estimate shape and texture that are of high-fidelity and faithful to input images. We also propose a novel technique to calculate more stable image-level gradients via diffusion models compared to existing alternatives. Finally, we produce realistic animations by fine-tuning the rendered shape and texture under rigid part transformations. Extensive evaluations on multiple existing datasets as well as newly introduced noisy web image collections with occlusions and truncation demonstrate that ARTIC3D outputs are more robust to noisy images, higher quality in terms of shape and texture details, and more realistic when animated.

In recent years, many researchers have proposed new models for synaptic plasticity in the brain based on principles of machine learning. The central motivation has been the development of learning algorithms that are able to learn difficult tasks while qualifying as "biologically plausible". However, the concept of a biologically plausible learning algorithm is only heuristically defined as an algorithm that is potentially implementable by biological neural networks. Further, claims that neural circuits could implement any given algorithm typically rest on an amorphous concept of "locality" (both in space and time). As a result, it is unclear what many proposed local learning algorithms actually predict biologically, and which of these are consequently good candidates for experimental investigation. Here, we address this lack of clarity by proposing formal and operational definitions of locality. Specifically, we define different classes of locality, each of which makes clear what quantities cannot be included in a learning rule if an algorithm is to qualify as local with respect to a given (biological) constraint. We subsequently use this framework to distill testable predictions from various classes of biologically plausible synaptic plasticity models that are robust to arbitrary choices about neural network architecture. Therefore, our framework can be used to guide claims of biological plausibility and to identify potential means of experimentally falsifying a proposed learning algorithm for the brain.

Federated Learning (FL) allows machine learning models to train locally on individual mobile devices, synchronizing model updates via a shared server. This approach safeguards user privacy; however, it also generates a heterogeneous training environment due to the varying performance capabilities across devices. As a result, “straggler” devices with lower performance often dictate the overall training time in FL. In this work, we aim to alleviate this performance bottleneck due to stragglers by dynamically balancing the training load across the system. We introduce Invariant Dropout, a method that extracts a sub-model based on the weight update threshold, thereby minimizing potential impacts on accuracy. Building on this dropout technique, we develop an adaptive training framework, Federated Learning using Invariant Dropout (FLuID). FLuID offers a lightweight sub-model extraction to regulate computational intensity, thereby reducing the load on straggler devices without affecting model quality. Our method leverages neuron updates from non-straggler devices to construct a tailored sub-model for each straggler based on client performance profiling. Furthermore, FLuID can dynamically adapt to changes in stragglers as runtime conditions shift. We evaluate FLuID using five real-world mobile clients. The evaluations show that Invariant Dropout maintains baseline model efficiency while alleviating the performance bottleneck of stragglers through a dynamic, runtime approach.

Most recurrent neural networks (RNNs) do not include a fundamental constraint of real neural circuits: Dale's Law, which implies that neurons must be excitatory (E) or inhibitory (I). Dale's Law is generally absent from RNNs because simply partitioning a standard network's units into E and I populations impairs learning. However, here we extend a recent feedforward bio-inspired EI network architecture, named Dale's ANNs, to recurrent networks, and demonstrate that good performance is possible while respecting Dale's Law. This begs the question: What makes some forms of EI network learn poorly and others learn well? And, why does the simple approach of incorporating Dale's Law impair learning? Historically the answer was thought to be the sign constraints on EI network parameters, and this was a motivation behind Dale's ANNs. However, here we show the spectral properties of the recurrent weight matrix at initialisation are more impactful on network performance than sign constraints. We find that simple EI partitioning results in a singular value distribution that is multimodal and dispersed, whereas standard RNNs have an unimodal, more clustered singular value distribution, as do recurrent Dale's ANNs. We also show that the spectral properties and performance of partitioned EI networks are worse for small networks with fewer I units, and we present normalised SVD entropy as a measure of spectrum pathology that correlates with performance. Overall, this work sheds light on a long-standing mystery in neuroscience-inspired AI and computational neuroscience, paving the way for greater alignment between neural networks and biology.

Molecular representation learning lays the foundation for drug discovery. However, existing methods suffer from poor out-of-distribution (OOD) generalization, particularly when data for training and testing originate from different environments. To address this issue, we propose a new framework for learning molecular representations that exhibit invariance and robustness against distribution shifts. Specifically, we propose a strategy called ``first-encoding-then-separation'' to identify invariant molecule features in the latent space, which deviates from conventional practices. Prior to the separation step, we introduce a residual vector quantization module that mitigates the over-fitting to training data distributions while preserving the expressivity of encoders. Furthermore, we design a task-agnostic self-supervised learning objective to encourage precise invariance identification, which enables our method widely applicable to a variety of tasks, such as regression and multi-label classification. Extensive experiments on 18 real-world molecular datasets demonstrate that our model achieves stronger generalization against state-of-the-art baselines in the presence of various distribution shifts. Our code is available at https://github.com/HICAI-ZJU/iMoLD.

The limited availability of annotations in small molecule datasets presents a challenge to machine learning models. To address this, one common strategy is to collaborate with additional auxiliary datasets. However, having more data does not always guarantee improvements. Negative transfer can occur when the knowledge in the target dataset differs or contradicts that of the auxiliary molecule datasets. In light of this, identifying the auxiliary molecule datasets that can benefit the target dataset when jointly trained remains a critical and unresolved problem. Through an empirical analysis, we observe that combining graph structure similarity and task similarity can serve as a more reliable indicator for identifying high-affinity auxiliary datasets. Motivated by this insight, we propose MolGroup, which separates the dataset affinity into task and structure affinity to predict the potential benefits of each auxiliary molecule dataset. MolGroup achieves this by utilizing a routing mechanism optimized through a bi-level optimization framework. Empowered by the meta gradient, the routing mechanism is optimized toward maximizing the target dataset's performance and quantifies the affinity as the gating score. As a result, MolGroup is capable of predicting the optimal combination of auxiliary datasets for each target dataset. Our extensive experiments demonstrate the efficiency and effectiveness of MolGroup, showing an average improvement of 4.41%/3.47% for GIN/Graphormer trained with the group of molecule datasets selected by MolGroup on 11 target molecule datasets.

Quantization is an effective way to compress neural networks. By reducing the bit width of the parameters, the processing efficiency of neural network models at edge devices can be notably improved. Most conventional quantization methods utilize real datasets to optimize quantization parameters and fine-tune. Due to the inevitable privacy and security issues of real samples, the existing real-data-driven methods are no longer applicable. Thus, a natural method is to introduce synthetic samples for zero-shot quantization (ZSQ). However, the conventional synthetic samples fail to retain the detailed texture feature distributions, which severely limits the knowledge transfer and performance of the quantized model. In this paper, a novel ZSQ method, TexQ is proposed to address this issue. We first synthesize a calibration image and extract its calibration center for each class with a texture feature energy distribution calibration method. Then, the calibration centers are used to guide the generator to synthesize samples. Finally, we introduce the mixup knowledge distillation module to diversify synthetic samples for fine-tuning. Extensive experiments on CIFAR10/100 and ImageNet show that TexQ is observed to perform state-of-the-art in ultra-low bit width quantization. For example, when ResNet-18 is quantized to 3-bit, TexQ achieves a 12.18% top-1 accuracy increase on ImageNet compared to state-of-the-art methods. Code at https://github.com/dangsingrue/TexQ.

We present a novel learning-based framework for reconstructing 3D structures from tilt-series cryo-Electron Tomography (cryo-ET) data. Cryo-ET is a powerful imaging technique that can achieve near-atomic resolutions. Still, it suffers from challenges such as missing-wedge acquisition, large data size, and high noise levels. Our framework addresses these challenges by using an adaptive tensorial-based representation for the 3D density field of the scanned sample. First, we optimize a quadtree structure to partition the volume of interest. Then, we learn a vector-matrix factorization of the tensor representing the density field in each node. Moreover, we use a loss function that combines a differentiable tomographic formation model with three regularization terms: total variation, boundary consistency constraint, and an isotropic Fourier prior. Our framework allows us to query the density at any location using the learned representation and obtain a high-quality 3D tomogram. We demonstrate the superiority of our framework over existing methods using synthetic and real data. Thus, our framework boosts the quality of the reconstruction while reducing the computation time and the memory footprint. The code is available at https://github.com/yuanhaowang1213/adaptivetensordf.

We study online meta-learning with bandit feedback, with the goal of improving performance across multiple tasks if they are similar according to some natural similarity measure. As the first to target the adversarial online-within-online partial-information setting, we design meta-algorithms that combine outer learners to simultaneously tune the initialization and other hyperparameters of an inner learner for two important cases: multi-armed bandits (MAB) and bandit linear optimization (BLO). For MAB, the meta-learners initialize and set hyperparameters of the Tsallis-entropy generalization of Exp3, with the task-averaged regret improving if the entropy of the optima-in-hindsight is small. For BLO, we learn to initialize and tune online mirror descent (OMD) with self-concordant barrier regularizers, showing that task-averaged regret varies directly with an action space-dependent measure they induce. Our guarantees rely on proving that unregularized follow-the-leader combined with two levels of low-dimensional hyperparameter tuning is enough to learn a sequence of affine functions of non-Lipschitz and sometimes non-convex Bregman divergences bounding the regret of OMD.

We introduce the Cambridge Law Corpus (CLC), a dataset for legal AI research. It consists of over 250 000 court cases from the UK. Most cases are from the 21st century, but the corpus includes cases as old as the 16th century. This paper presents the first release of the corpus, containing the raw text and meta-data. Together with the corpus, we provide annotations on case outcomes for 638 cases, done by legal experts. Using our annotated data, we have trained and evaluated case outcome extraction with GPT-3, GPT-4 and RoBERTa models to provide benchmarks. We include an extensive legal and ethical discussion to address the potentially sensitive nature of this material. As a consequence, the corpus will only be released for research purposes under certain restrictions.