NeurIPS 2023 Papers

Deep Neural Networks (DNNs) have been criticized because they easily overfit noisy (incorrect) labels. To improve the robustness of DNNs, existing methods for image data regard samples with small training losses as correctly labeled data (small-loss criterion). Nevertheless, time series' discriminative patterns are easily distorted by external noises (i.e., frequency perturbations) during the recording process. This results in training losses of some time series samples that do not meet the small-loss criterion. Therefore, this paper proposes a deep learning paradigm called Scale-teaching to cope with time series noisy labels. Specifically, we design a fine-to-coarse cross-scale fusion mechanism for learning discriminative patterns by utilizing time series at different scales to train multiple DNNs simultaneously. Meanwhile, each network is trained in a cross-teaching manner by using complementary information from different scales to select small-loss samples as clean labels. For unselected large-loss samples, we introduce multi-scale embedding graph learning via label propagation to correct their labels by using selected clean samples. Experiments on multiple benchmark time series datasets demonstrate the superiority of the proposed Scale-teaching paradigm over state-of-the-art methods in terms of effectiveness and robustness.

Training on large-scale datasets can boost the performance of video instance segmentation while the annotated datasets for VIS are hard to scale up due to the high labor cost. What we possess are numerous isolated filed-specific datasets, thus, it is appealing to jointly train models across the aggregation of datasets to enhance data volume and diversity. However, due to the heterogeneity in category space, as mask precision increase with the data volume, simply utilizing multiple datasets will dilute the attention of models on different taxonomy. Thus, increasing the data scale and enriching taxonomy space while improving classification precision is important. In this work, we analyze that providing extra taxonomy information can help models concentrate on specific taxonomy, and propose our model named Taxonomy-aware Multi-dataset Joint Training for Video Instance Segmentation (TMT-VIS) to address this vital challenge. Specifically, we design a two-stage taxonomy aggregation module that first compiles taxonomy information from input videos and then aggregates these taxonomy priors into instance queries before the transformer decoder. We conduct extensive experimental evaluations on four popular and challenging benchmarks, including YouTube-VIS 2019, YouTube-VIS 2021, OVIS, and UVO. Our model shows significant improvement over the baseline solutions, and sets new state-of-the-art records on all these benchmarks. These appealing and encouraging results demonstrate the effectiveness and generality of our proposed approach. The code and trained models will be publicly available.

The power of DNNs relies heavily on the quantity and quality of training data. However, collecting and annotating data on a large scale is often expensive and time-consuming. To address this issue, we explore a new task, termed dataset expansion, aimed at expanding a ready-to-use small dataset by automatically creating new labeled samples. To this end, we present a Guided Imagination Framework (GIF) that leverages cutting-edge generative models like DALL-E2 and Stable Diffusion (SD) to "imagine" and create informative new data from the input seed data. Specifically, GIF conducts data imagination by optimizing the latent features of the seed data in the semantically meaningful space of the prior model, resulting in the creation of photo-realistic images with new content. To guide the imagination towards creating informative samples for model training, we introduce two key criteria, i.e., class-maintained information boosting and sample diversity promotion. These criteria are verified to be essential for effective dataset expansion: GIF-SD obtains 13.5% higher model accuracy on natural image datasets than unguided expansion with SD. With these essential criteria, GIF successfully expands small datasets in various scenarios, boosting model accuracy by 36.9% on average over six natural image datasets and by 13.5% on average over three medical datasets. The source code is available at https://github.com/Vanint/DatasetExpansion.

We study the mean estimation problem under communication and local differential privacy constraints. While previous work has proposed order-optimal algorithms for the same problem (i.e., asymptotically optimal as we spend more bits), exact optimality (in the non-asymptotic setting) still has not been achieved. In this work, we take a step towards characterizing the exact-optimal approach in the presence of shared randomness (a random variable shared between the server and the user) and identify several conditions for exact optimality. We prove that one of the conditions is to utilize a rotationally symmetric shared random codebook. Based on this, we propose a randomization mechanism where the codebook is a randomly rotated simplex -- satisfying the properties of the exact-optimal codebook. The proposed mechanism is based on a $k$-closest encoding which we prove to be exact-optimal for the randomly rotated simplex codebook.

Data-Free Model Extraction (DFME) aims to clone a black-box model without knowing its original training data distribution, making it much easier for attackers to steal commercial models. Defense against DFME faces several challenges: (i) effectiveness; (ii) efficiency; (iii) no prior on the attacker's query data distribution and strategy. However, existing defense methods: (1) are highly computation and memory inefficient; or (2) need strong assumptions about attack data distribution; or (3) can only delay the attack or prove a model theft after the model stealing has happened. In this work, we propose a Memory and Computation efficient defense approach, named MeCo, to prevent DFME from happening while maintaining the model utility simultaneously by distributionally robust defensive training on the target victim model. Specifically, we randomize the input so that it: (1) causes a mismatch of the knowledge distillation loss for attackers; (2) disturbs the zeroth-order gradient estimation; (3) changes the label prediction for the attack query data. Therefore, the attacker can only extract misleading information from the black-box model. Extensive experiments on defending against both decision-based and score-based DFME demonstrate that MeCo can significantly reduce the effectiveness of existing DFME methods and substantially improve running efficiency.

We study the distributed tracking model, also known as distributed functional monitoring. This model involves $k$ sites each receiving a stream of items and communicating with the central server. The server's task is to track a function of all items received thus far continuously, with minimum communication cost. For count tracking, it is known that there is a $\sqrt{k}$ gap in communication between deterministic and randomized algorithms. However, existing randomized algorithms assume an "oblivious adversary" who constructs the entire input streams before the algorithm starts. Here we consider adaptive adversaries who can choose new items based on previous answers from the algorithm. Deterministic algorithms are trivially robust to adaptive adversaries, while randomized ones may not. Therefore, we investigate whether the $\sqrt{k}$ advantage of randomized algorithms is from randomness itself or the oblivious adversary assumption. We provide an affirmative answer to this question by giving a robust algorithm with optimal communication. Existing robustification techniques do not yield optimal bounds due to the inherent challenges of the distributed nature of the problem. To address this, we extend the differential privacy framework by introducing "partial differential privacy" and proving a new generalization theorem. This theorem may have broader applications beyond robust count tracking, making it of independent interest.

Large language models (LLMs) have been recently leveraged as training data generators for various natural language processing (NLP) tasks. While previous research has explored different approaches to training models using generated data, they generally rely on simple class-conditional prompts, which may limit the diversity of the generated data and inherit systematic biases of LLM. Thus, we investigate training data generation with diversely attributed prompts (e.g., specifying attributes like length and style), which have the potential to yield diverse and attributed generated data. Our investigation focuses on datasets with high cardinality and diverse domains, wherein we demonstrate that attributed prompts outperform simple class-conditional prompts in terms of the resulting model's performance. Additionally, we present a comprehensive empirical study on data generation encompassing vital aspects like bias, diversity, and efficiency, and highlight three key observations: firstly, synthetic datasets generated by simple prompts exhibit significant biases, such as regional bias; secondly, attribute diversity plays a pivotal role in enhancing model performance; lastly, attributed prompts achieve the performance of simple class-conditional prompts while utilizing only 5\% of the querying cost of ChatGPT associated with the latter. The data and code are available on {\url{https://github.com/yueyu1030/AttrPrompt}}.

Weight-sharing quantization has emerged as a technique to reduce energy expenditure during inference in large neural networks by constraining their weights to a limited set of values. However, existing methods often assume weights are treated solely based on value, neglecting the unique role of weight position. This paper proposes a probabilistic framework based on Bayesian neural networks (BNNs) and a variational relaxation to identify which weights can be moved to which cluster center and to what degree based on their individual position-specific learned uncertainty distributions. We introduce a new initialization setting and a regularization term, enabling the training of BNNs with complex dataset-model combinations. Leveraging the flexibility of weight values from probability distributions, we enhance noise resilience and compressibility. Our iterative clustering procedure demonstrates superior compressibility and higher accuracy compared to state-of-the-art methods on both ResNet models and the more complex transformer-based architectures. In particular, our method outperforms the state-of-the-art quantization method top-1 accuracy by 1.6\% on ImageNet using DeiT-Tiny, with its 5 million+ weights now represented by only 296 unique values. Code available at https://github.com/subiawaud/PWFN.

The tremendous success of large models trained on extensive datasets demonstrates that scale is a key ingredient in achieving superior results. Therefore, the reflection on the rationality of designing knowledge distillation (KD) approaches for limited-capacity architectures solely based on small-scale datasets is now deemed imperative. In this paper, we identify the small data pitfall that presents in previous KD methods, which results in the underestimation of the power of vanilla KD framework on large-scale datasets such as ImageNet-1K. Specifically, we show that employing stronger data augmentation techniques and using larger datasets can directly decrease the gap between vanilla KD and other meticulously designed KD variants. This highlights the necessity of designing and evaluating KD approaches in the context of practical scenarios, casting off the limitations of small-scale datasets. Our investigation of the vanilla KD and its variants in more complex schemes, including stronger training strategies and different model capacities, demonstrates that vanilla KD is elegantly simple but astonishingly effective in large-scale scenarios. Without bells and whistles, we obtain state-of-the-art ResNet-50, ViT-S, and ConvNeXtV2-T models for ImageNet, which achieve 83.1%, 84.3%, and 85.0% top-1 accuracy, respectively. PyTorch code and checkpoints can be found at https://github.com/Hao840/vanillaKD.

Mislabeled, duplicated, or biased data in real-world scenarios can lead to prolonged training and even hinder model convergence. Traditional solutions prioritizing easy or hard samples lack the flexibility to handle such a variety simultaneously. Recent work has proposed a more reasonable data selection principle by examining the data's impact on the model's generalization loss. However, its practical adoption relies on less principled approximations and additional holdout data. This work solves these problems by leveraging a lightweight Bayesian treatment and incorporating off-the-shelf zero-shot predictors built on large-scale pre-trained models. The resulting algorithm is efficient and easy to implement. We perform extensive empirical studies on challenging benchmarks with considerable data noise and imbalance in the online batch selection scenario, and observe superior training efficiency over competitive baselines. Notably, on the challenging WebVision benchmark, our method can achieve similar predictive performance with significantly fewer training iterations than leading data selection methods.

Langevin Monte Carlo (LMC) and its stochastic gradient versions are powerful algorithms for sampling from complex high-dimensional distributions. To sample from a distribution with density $\pi(\theta)\propto \exp(-U(\theta)) $, LMC iteratively generates the next sample by taking a step in the gradient direction $\nabla U$ with added Gaussian perturbations. Expectations w.r.t. the target distribution $\pi$ are estimated by averaging over LMC samples. In ordinary Monte Carlo, it is well known that the estimation error can be substantially reduced by replacing independent random samples by quasi-random samples like low-discrepancy sequences. In this work, we show that the estimation error of LMC can also be reduced by using quasi-random samples. Specifically, we propose to use completely uniformly distributed (CUD) sequences with certain low-discrepancy property to generate the Gaussian perturbations. Under smoothness and convexity conditions, we prove that LMC with a low-discrepancy CUD sequence achieves smaller error than standard LMC. The theoretical analysis is supported by compelling numerical experiments, which demonstrate the effectiveness of our approach.

This work is concerned with the estimation problem of linear model when the sample size is extremely large and the data dimension can vary with the sample size. In this setting, the least square estimator based on the full data is not feasible with limited computational resources. Many existing methods for this problem are based on the sketching technique which uses the sketched data to perform least square estimation. We derive fine-grained lower bounds of the conditional mean squared error for sketching methods. For sampling methods, our lower bound provides an attainable optimal convergence rate. Our result implies that when the dimension is large, there is hardly a sampling method can have a faster convergence rate than the uniform sampling method. To achieve a better statistical performance, we propose a new sketching method based on data averaging. The proposed method reduces the original data to a few averaged observations. These averaged observations still satisfy the linear model and are used to estimate the regression coefficients. The asymptotic behavior of the proposed estimation procedure is studied. Our theoretical results show that the proposed method can achieve a faster convergence rate than the optimal convergence rate for sampling methods. Theoretical and numerical results show that the proposed estimator has good statistical performance as well as low computational cost.

Neural network pruning and quantization techniques are almost as old as neural networks themselves. However, to date, only ad-hoc comparisons between the two have been published. In this paper, we set out to answer the question of which is better: neural network quantization or pruning? By answering this question, we hope to inform design decisions made on neural network hardware going forward. We provide an extensive comparison between the two techniques for compressing deep neural networks. First, we give an analytical comparison of expected quantization and pruning error for general data distributions. Then, we provide lower and upper bounds for the per-layer pruning and quantization error in trained networks and compare these to empirical error after optimization. Finally, we provide an extensive experimental comparison for training 8 large-scale models trained on 3 tasks and provide insights into the representations learned during fine-tuning with quantization and pruning in the loop. Our results show that in most cases quantization outperforms pruning. Only in some scenarios with a very high compression ratio, compression might be beneficial from an accuracy standpoint.