NeurIPS 2022 论文列表

3D-aware image synthesis aims at learning a generative model that can render photo-realistic 2D images while capturing decent underlying 3D shapes. A popular solution is to adopt the generative adversarial network (GAN) and replace the generator with a 3D renderer, where volume rendering with neural radiance field (NeRF) is commonly used. Despite the advancement of synthesis quality, existing methods fail to obtain moderate 3D shapes. We argue that, considering the two-player game in the formulation of GANs, only making the generator 3D-aware is not enough. In other words, displacing the generative mechanism only offers the capability, but not the guarantee, of producing 3D-aware images, because the supervision of the generator primarily comes from the discriminator. To address this issue, we propose GeoD through learning a geometry-aware discriminator to improve 3D-aware GANs. Concretely, besides differentiating real and fake samples from the 2D image space, the discriminator is additionally asked to derive the geometry information from the inputs, which is then applied as the guidance of the generator. Such a simple yet effective design facilitates learning substantially more accurate 3D shapes. Extensive experiments on various generator architectures and training datasets verify the superiority of GeoD over state-of-the-art alternatives. Moreover, our approach is registered as a general framework such that a more capable discriminator (i.e., with a third task of novel view synthesis beyond domain classification and geometry extraction) can further assist the generator with a better multi-view consistency. Project page can be found at https://vivianszf.github.io/geod.

Generalization and invariance are two essential properties of machine learning models. Generalization captures a model's ability to classify unseen data while invariance measures consistency of model predictions on transformations of the data. Existing research suggests a positive relationship: a model generalizing well should be invariant to certain visual factors. Building on this qualitative implication we make two contributions. First, we introduce effective invariance (EI), a simple and reasonable measure of model invariance which does not rely on image labels. Given predictions on a test image and its transformed version, EI measures how well the predictions agree and with what level of confidence. Second, using invariance scores computed by EI, we perform large-scale quantitative correlation studies between generalization and invariance, focusing on rotation and grayscale transformations. From a model-centric view, we observe generalization and invariance of different models exhibit a strong linear relationship, on both in-distribution and out-of-distribution datasets. From a dataset-centric view, we find a certain model's accuracy and invariance linearly correlated on different test sets. Apart from these major findings, other minor but interesting insights are also discussed.

Model attributions are important in deep neural networks as they aid practitioners in understanding the models, but recent studies reveal that attributions can be easily perturbed by adding imperceptible noise to the input. The non-differentiable Kendall's rank correlation is a key performance index for attribution protection. In this paper, we first show that the expected Kendall's rank correlation is positively correlated to cosine similarity and then indicate that the direction of attribution is the key to attribution robustness. Based on these findings, we explore the vector space of attribution to explain the shortcomings of attribution defense methods using $\ell_p$ norm and propose integrated gradient regularizer (IGR), which maximizes the cosine similarity between natural and perturbed attributions. Our analysis further exposes that IGR encourages neurons with the same activation states for natural samples and the corresponding perturbed samples. Our experiments on different models and datasets confirm our analysis on attribution protection and demonstrate a decent improvement in adversarial robustness.

We give a simple, generic conformal prediction method for sequential prediction that achieves target empirical coverage guarantees on adversarial data. It is computationally lightweight --- comparable to split conformal prediction --- but does not require having a held-out validation set, and so all data can be used for training models from which to derive a conformal score. Furthermore, it gives stronger than marginal coverage guarantees in two ways. First, it gives threshold-calibrated prediction sets that have correct empirical coverage even conditional on the threshold used to form the prediction set from the conformal score. Second, the user can specify an arbitrary collection of subsets of the feature space --- possibly intersecting --- and the coverage guarantees will also hold conditional on membership in each of these subsets. We call our algorithm MVP, short for MultiValid Prediction. We give both theory and an extensive set of empirical evaluations.

We study the densest subgraph problem (DSG) and the densest subgraph local decomposition problem (DSG-LD) in undirected graphs. We also consider supermodular generalizations of these problems. For large scale graphs simple iterative algorithms perform much better in practice than theoretically fast algorithms based on network-flow or LP solvers. Boob et al [1] recently gave a fast iterative algorithm called Greedy++ for DSG. It was shown in [2] that it converges to a $(1-\epsilon)$ relative approximation to the optimum density in $O(\frac{1}{\epsilon^2} \frac{\Delta(G)}{\lambda^*})$ iterations where $\Delta(G)$ is the maximum degree and $\lambda^*$ is the optimum density. Danisch et al. [3] gave an iterative algorithm based on the Frank-Wolfe algorithm for DSG-LD that takes $O(\frac{m\Delta(G) }{\epsilon^2})$ iterations to converge to an $\epsilon$-additive approximate local decomposition vector $\hat{b}$, where $m$ is number of edges in the graph. In this paper we give a new iterative algorithm for both problems that takes at most $O(\frac{\sqrt{m\Delta(G)}}{\epsilon})$ iterations to converge to an $\epsilon$-additive approximate local decomposition vector; each iteration can be implemented in $O(m)$ time. We describe a fractional peeling technique which has strong empirical performance as well as theoretical guarantees. The algorithm is scalable and simple, and can be applied to graphs with hundreds of millions of edges. We test our algorithm on real and synthetic data sets and show that it provides a significant benefit over previous algorithms. The algorithm and analysis extends to hypergraphs.

A striking observation about iterative magnitude pruning (IMP; Frankle et al. 2020) is that—after just a few hundred steps of dense training—the method can find a sparse sub-network that can be trained to the same accuracy as the dense network. However, the same does not hold at step 0, i.e. random initialization. In this work, we seek to understand how this early phase of pre-training leads to a good initialization for IMP both through the lens of the data distribution and the loss landscape geometry. Empirically we observe that, holding the number of pre-training iterations constant, training on a small fraction of (randomly chosen) data suffices to obtain an equally good initialization for IMP. We additionally observe that by pre-training only on "easy" training data, we can decrease the number of steps necessary to find a good initialization for IMP compared to training on the full dataset or a randomly chosen subset. Finally, we identify novel properties of the loss landscape of dense networks that are predictive of IMP performance, showing in particular that more examples being linearly mode connected in the dense network correlates well with good initializations for IMP. Combined, these results provide new insight into the role played by the early phase training in IMP.

Previous works show global covariance pooling (GCP) has great potential to improve deep architectures especially on visual recognition tasks, where post-normalization of GCP plays a very important role in final performance. Although several post-normalization strategies have been studied, these methods pay more close attention to effect of normalization on covariance representations rather than the whole GCP networks, and their effectiveness requires further understanding. Meanwhile, existing effective post-normalization strategies (e.g., matrix power normalization) usually suffer from high computational complexity (e.g., $O(d^{3})$ for $d$-dimensional inputs). To handle above issues, this work first analyzes the effect of post-normalization from the perspective of training GCP networks. Particularly, we for the first time show that \textit{effective post-normalization can make a good trade-off between representation decorrelation and information preservation for GCP, which are crucial to alleviate over-fitting and increase representation ability of deep GCP networks, respectively}. Based on this finding, we can improve existing post-normalization methods with some small modifications, providing further support to our observation. Furthermore, this finding encourages us to propose a novel pre-normalization method for GCP (namely DropCov), which develops an adaptive channel dropout on features right before GCP, aiming to reach trade-off between representation decorrelation and information preservation in a more efficient way. Our DropCov only has a linear complexity of $O(d)$, while being free for inference. Extensive experiments on various benchmarks (i.e., ImageNet-1K, ImageNet-C, ImageNet-A, Stylized-ImageNet, and iNat2017) show our DropCov is superior to the counterparts in terms of efficiency and effectiveness, and provides a simple yet effective method to improve performance of deep architectures involving both deep convolutional neural networks (CNNs) and vision transformers (ViTs).

We investigate the generalization capabilities of neural signed distance functions (SDFs) for learning 3D object representations for unseen and unlabeled point clouds. Existing methods can fit SDFs to a handful of object classes and boast fine detail or fast inference speeds, but do not generalize well to unseen shapes. We introduce a two-stage semi-supervised meta-learning approach that transfers shape priors from labeled to unlabeled data to reconstruct unseen object categories. The first stage uses an episodic training scheme to simulate training on unlabeled data and meta-learns initial shape priors. The second stage then introduces unlabeled data with disjoint classes in a semi-supervised scheme to diversify these priors and achieve generalization. We assess our method on both synthetic data and real collected point clouds. Experimental results and analysis validate that our approach outperforms existing neural SDF methods and is capable of robust zero-shot inference on 100+ unseen classes. Code can be found at https://github.com/princeton-computational-imaging/gensdf

One-shot Federated Learning (FL) has recently emerged as a promising approach, which allows the central server to learn a model in a single communication round. Despite the low communication cost, existing one-shot FL methods are mostly impractical or face inherent limitations, \eg a public dataset is required, clients' models are homogeneous, and additional data/model information need to be uploaded. To overcome these issues, we propose a novel two-stage \textbf{D}ata-fre\textbf{E} o\textbf{N}e-\textbf{S}hot federated l\textbf{E}arning (DENSE) framework, which trains the global model by a data generation stage and a model distillation stage. DENSE is a practical one-shot FL method that can be applied in reality due to the following advantages: (1) DENSE requires no additional information compared with other methods (except the model parameters) to be transferred between clients and the server; (2) DENSE does not require any auxiliary dataset for training; (3) DENSE considers model heterogeneity in FL, \ie different clients can have different model architectures. Experiments on a variety of real-world datasets demonstrate the superiority of our method. For example, DENSE outperforms the best baseline method Fed-ADI by 5.08\% on CIFAR10 dataset.

Training deep neural network classifiers that are certifiably robust against adversarial attacks is critical to ensuring the security and reliability of AI-controlled systems. Although numerous state-of-the-art certified training methods have been developed, they are computationally expensive and scale poorly with respect to both dataset and network complexity. Widespread usage of certified training is further hindered by the fact that periodic retraining is necessary to incorporate new data and network improvements. In this paper, we propose Certified Robustness Transfer (CRT), a general-purpose framework for reducing the computational overhead of any certifiably robust training method through knowledge transfer. Given a robust teacher, our framework uses a novel training loss to transfer the teacher’s robustness to the student. We provide theoretical and empirical validation of CRT. Our experiments on CIFAR-10 show that CRT speeds up certified robustness training by 8× on average across three different architecture generations while achieving comparable robustness to state-of-the-art methods. We also show that CRT can scale to large-scale datasets like ImageNet.

We study linear regression from data distributed over a network of agents (with no master node) under high-dimensional scaling, which allows the ambient dimension to grow faster than the sample size. We propose a novel decentralization of the projected gradient algorithm whereby agents iteratively update their local estimates by a “double-mixing” mechanism, which suitably combines averages of iterates and gradients of neighbouring nodes. Under standard assumptions on the statistical model and network connectivity, the proposed method enjoys global linear convergence up to the statistical precision of the model. This improves on guarantees of (plain) DGD algorithms, whose iteration complexity grows undesirably with the ambient dimension. Our technical contribution is a novel convergence analysis that resembles (albeit different) algorithmic stability arguments extended to high-dimensions and distributed setting, which is of independent interest.

Communication compression is a crucial technique for modern distributed learning systems to alleviate their communication bottlenecks over slower networks. Despite recent intensive studies of gradient compression for data parallel-style training, compressing the activations for models trained with pipeline parallelism is still an open problem. In this paper, we propose AQ-SGD, a novel activation compression algorithm for communication-efficient pipeline parallelism training over slow networks. Different from previous efforts in activation compression, instead of compressing activation values directly, AQ-SGD compresses the changes of the activations. This allows us to show, to the best of our knowledge for the first time, that one can still achieve $O(1/\sqrt{T})$ convergence rate for non-convex objectives under activation compression, without making assumptions on gradient unbiasedness that do not hold for deep learning models with non-linear activation functions. We then show that AQ-SGD can be optimized and implemented efficiently, without additional end-to-end runtime overhead. We evaluated AQ-SGD to fine-tune language models with up to 1.5 billion parameters, compressing activation to 2-4 bits. AQ-SGD provides up to $4.3\times$ end-to-end speed-up in slower networks, without sacrificing model quality. Moreover, we also show that AQ-SGD can be combined with state-of-the-art gradient compression algorithms to enable end-to-end communication compression: All communications between machines, including model gradients, forward activations, and backward gradients are compressed into lower precision. This provides up to $4.9\times$ end-to-end speed-up, without sacrificing model quality.

Multiple time series data occur in many real applications and the alignment among them is usually a fundamental step of data analysis. Frequently, these multiple time series are inter-dependent, which provides extra information for the alignment task and this information cannot be well utilized in the conventional pairwise alignment methods. Recently, the joint alignment was modeled as a max-flow problem, in which both the profile similarity between the aligned time series and the distance between adjacent warping functions are jointly optimized. However, despite the new model having elegant mathematical formulation and superior alignment accuracy, the long computation time and large memory usage, due to the use of the existing general-purpose max-flow algorithms, limit significantly its well-deserved wide use. In this report, we present BIdirectional pushing with Linear Component Operations (BILCO), a novel algorithm that solves the joint alignment max-flow problems efficiently and exactly. We develop the strategy of linear component operations that integrates dynamic programming technique and the push-relabel approach. This strategy is motivated by the fact that the joint alignment max-flow problem is a generalization of dynamic time warping (DTW) and numerous individual DTW problems are embedded. Further, a bidirectional-pushing strategy is proposed to introduce prior knowledge and reduce unnecessary computation, by leveraging another fact that good initialization can be easily computed for the joint alignment max-flow problem. We demonstrate the efficiency of BILCO using both synthetic and real experiments. Tested on thousands of datasets under various simulated scenarios and in three distinct application categories, BILCO consistently achieves at least 10 and averagely 20-folds increase in speed, and uses at most 1/8 and averagely 1/10 memory compared with the best existing max-flow method. Our source code can be found at https://github.com/yu-lab-vt/BILCO.

Knowledge distillation(KD) is a widely-used technique to train compact models in object detection. However, there is still a lack of study on how to distill between heterogeneous detectors. In this paper, we empirically find that better FPN features from a heterogeneous teacher detector can help the student although their detection heads and label assignments are different. However, directly aligning the feature maps to distill detectors suffers from two problems. First, the difference in feature magnitude between the teacher and the student could enforce overly strict constraints on the student. Second, the FPN stages and channels with large feature magnitude from the teacher model could dominate the gradient of distillation loss, which will overwhelm the effects of other features in KD and introduce much noise. To address the above issues, we propose to imitate features with Pearson Correlation Coefficient to focus on the relational information from the teacher and relax constraints on the magnitude of the features. Our method consistently outperforms the existing detection KD methods and works for both homogeneous and heterogeneous student-teacher pairs. Furthermore, it converges faster. With a powerful MaskRCNN-Swin detector as the teacher, ResNet-50 based RetinaNet and FCOS achieve 41.5% and 43.9% $mAP$ on COCO2017, which are 4.1% and 4.8% higher than the baseline, respectively.

Prompt learning approaches have made waves in natural language processing by inducing better few-shot performance while they still follow a parametric-based learning paradigm; the oblivion and rote memorization problems in learning may encounter unstable generalization issues. Specifically, vanilla prompt learning may struggle to utilize atypical instances by rote during fully-supervised training or overfit shallow patterns with low-shot data. To alleviate such limitations, we develop RetroPrompt with the motivation of decoupling knowledge from memorization to help the model strike a balance between generalization and memorization. In contrast with vanilla prompt learning, RetroPrompt constructs an open-book knowledge-store from training instances and implements a retrieval mechanism during the process of input, training and inference, thus equipping the model with the ability to retrieve related contexts from the training corpus as cues for enhancement. Extensive experiments demonstrate that RetroPrompt can obtain better performance in both few-shot and zero-shot settings. Besides, we further illustrate that our proposed RetroPrompt can yield better generalization abilities with new datasets. Detailed analysis of memorization indeed reveals RetroPrompt can reduce the reliance of language models on memorization; thus, improving generalization for downstream tasks. Code is available in https://github.com/zjunlp/PromptKG/tree/main/research/RetroPrompt.

In this paper, we focus on multi-task classification, where related classification tasks share the same label space and are learned simultaneously. In particular, we tackle a new setting, which is more realistic than currently addressed in the literature, where categories shift from training to test data. Hence, individual tasks do not contain complete training data for the categories in the test set. To generalize to such test data, it is crucial for individual tasks to leverage knowledge from related tasks. To this end, we propose learning an association graph to transfer knowledge among tasks for missing classes. We construct the association graph with nodes representing tasks, classes and instances, and encode the relationships among the nodes in the edges to guide their mutual knowledge transfer. By message passing on the association graph, our model enhances the categorical information of each instance, making it more discriminative. To avoid spurious correlations between task and class nodes in the graph, we introduce an assignment entropy maximization that encourages each class node to balance its edge weights. This enables all tasks to fully utilize the categorical information from related tasks. An extensive evaluation on three general benchmarks and a medical dataset for skin lesion classification reveals that our method consistently performs better than representative baselines.

Tensor decompositions are powerful tools for dimensionality reduction and feature interpretation of multidimensional data such as signals. Existing tensor decomposition objectives (e.g., Frobenius norm) are designed for fitting raw data under statistical assumptions, which may not align with downstream classification tasks. In practice, raw input tensor can contain irrelevant information while data augmentation techniques may be used to smooth out class-irrelevant noise in samples. This paper addresses the above challenges by proposing augmented tensor decomposition (ATD), which effectively incorporates data augmentations and self-supervised learning (SSL) to boost downstream classification. To address the non-convexity of the new augmented objective, we develop an iterative method that enables the optimization to follow an alternating least squares (ALS) fashion. We evaluate our proposed ATD on multiple datasets. It can achieve 0.8%~2.5% accuracy gain over tensor-based baselines. Also, our ATD model shows comparable or better performance (e.g., up to 15% in accuracy) over self-supervised and autoencoder baselines while using less than 5% of learnable parameters of these baseline models.

There is currently a large gap in performance between the statistically rigorous methods like linear regression or additive splines and the powerful deep methods using neural networks. Previous works attempting to close this gap have failed to fully consider the exponentially growing number of feature combinations which deep networks consider automatically during training. In this work, we develop a tractable selection algorithm to efficiently identify the necessary feature combinations by leveraging techniques in feature interaction detection. Our proposed Sparse Interaction Additive Networks (SIAN) construct a bridge from these simple and interpretable models to a fully connected neural network. SIAN achieves competitive performance against state-of-the-art methods across multiple large-scale tabular datasets and consistently finds an optimal tradeoff between the modeling capacity of neural networks and the generalizability of simpler methods.

Weight sparsity is a promising approach to reducing the model size and computation cost of convolutional neural networks (CNNs). Nevertheless, non-zero weights often distribute randomly in sparse CNN models, introducing enormous difficulty in obtaining actual speedup on common hardware (e.g., GPU) over their dense counterparts. Existing acceleration solutions either require hardware modifications for irregular memory access support or rely on a partially structured sparsity pattern. Neither of these methods is capable of achieving fruitful speedup on convolution layers. In this work, we propose an algorithm-software co-designed sparse convolution based on a novel out-vector-wise (OVW) sparse pattern. Building on the insight that vertical vector integrity can preserve continuous memory access in IM2COL, the OVW pattern treats a $V\times1$ vector as an entirety. To reduce the error caused by sparsity, we propose an equivalent transformation process, i.e., clustering-based channel permutation, to gather similar rows together. Experimental evaluations demonstrate that our method achieves a $1.7\times$ and $3.2\times$ speedup over the SOTA solution and the dense convolution of ResNet50 on NVIDIA V100 at 75\% sparsity, respectively, with only negligible accuracy loss. Moreover, compared to the SOTA solution that achieves speedups only on data with 60\% sparsity or more, our method begins to obtain speedups on data with only 10\% sparsity.