NeurIPS 2023 论文列表

A compelling use case of offline reinforcement learning (RL) is to obtain a policy initialization from existing datasets followed by fast online fine-tuning with limited interaction. However, existing offline RL methods tend to behave poorly during fine-tuning. In this paper, we devise an approach for learning an effective initialization from offline data that also enables fast online fine-tuning capabilities. Our approach, calibrated Q-learning (Cal-QL), accomplishes this by learning a conservative value function initialization that underestimates the value of the learned policy from offline data, while also being calibrated, in the sense that the learned Q-values are at a reasonable scale. We refer to this property as calibration, and define it formally as providing a lower bound on the true value function of the learned policy and an upper bound on the value of some other (suboptimal) reference policy, which may simply be the behavior policy. We show that offline RL algorithms that learn such calibrated value functions lead to effective online fine-tuning, enabling us to take the benefits of offline initializations in online fine-tuning. In practice, Cal-QL can be implemented on top of the conservative Q learning (CQL) for offline RL within a one-line code change. Empirically, Cal-QL outperforms state-of-the-art methods on 9/11 fine-tuning benchmark tasks that we study in this paper. Code and video are available at https://nakamotoo.github.io/Cal-QL

The recent emergence of Large Language Models based on the Transformer architecture has enabled dramatic advancements in the field of Natural Language Processing. However, these models have long inference latency, which limits their deployment and makes them prohibitively expensive for various real-time applications. The inference latency is further exacerbated by autoregressive generative tasks, as models need to run iteratively to generate tokens sequentially without leveraging token-level parallelization. To address this, we propose Big Little Decoder (BiLD), a framework that can improve inference efficiency and latency for a wide range of text generation applications. The BiLD framework contains two models with different sizes that collaboratively generate text. The small model runs autoregressively to generate text with a low inference cost, and the large model is only invoked occasionally to refine the small model’s inaccurate predictions in a non-autoregressive manner. To coordinate the small and large models, BiLD introduces two simple yet effective policies: (1) the fallback policy that determines when to hand control over to the large model; and (2) the rollback policy that determines when the large model needs to correct the small model's inaccurate predictions. To evaluate our framework across different tasks and models, we apply BiLD to various text generation scenarios encompassing machine translation on IWSLT 2017 De-En and WMT 2014 De-En, and summarization on XSUM and CNN/DailyMail. On an NVIDIA T4 GPU, our framework achieves a speedup of up to 2.12x speedup with minimal generation quality degradation. Furthermore, our framework is fully plug-and-play and can be applied without any modifications in the training process or model architecture. Our code is open-sourced.

In recent years, big models based on Transformers have achieved state-of-the-art performance on many artificial intelligence (AI) tasks. Despite the success of these Transformer-based models, their huge parameter size poses a serious challenge to their training, both from the storage and computation perspectives. To this end, memory optimization (e.g., rematerialization and offloading) and parallelism (e.g., data parallelism and model parallelism) are widely explored to make training Transformers more efficient. In this paper, we propose a framework to automatically find an efficient integration of memory optimization and parallelism for High-Throughput Transformer Training (named H3T), which is rarely considered by existing efforts for training big Transformer-based models. Specifically, we design search algorithms to combine appropriate memory optimization strategies and parallelism schemes to achieve a balance between memory overhead and training efficiency. We implement H3T based on an open-source toolkit BMTrain and then use H3T to train the Transformers of different sizes to evaluate the efficiency of H3T. The experimental results show that H3T outperforms the most popular deep learning (DL) toolkit Megatron-DeepSpeed by $1.2\times \sim 4.3\times$ training speed while reducing $34.6\% \sim 80.5\%$ of memory overhead. Moreover, H3T can use only 64 NVIDIA A100 GPUs to train GPT-3-175B, which is very difficult for existing DL toolkits. The source code is available at https://github.com/OpenBMB/BMTrain/tree/h3t.

Multi-agent collaborative perception has recently received widespread attention as an emerging application in driving scenarios. Despite the advancements in previous efforts, challenges remain due to various noises in the perception procedure, including communication redundancy, transmission delay, and collaboration heterogeneity. To tackle these issues, we propose \textit{How2comm}, a collaborative perception framework that seeks a trade-off between perception performance and communication bandwidth. Our novelties lie in three aspects. First, we devise a mutual information-aware communication mechanism to maximally sustain the informative features shared by collaborators. The spatial-channel filtering is adopted to perform effective feature sparsification for efficient communication. Second, we present a flow-guided delay compensation strategy to predict future characteristics from collaborators and eliminate feature misalignment due to temporal asynchrony. Ultimately, a pragmatic collaboration transformer is introduced to integrate holistic spatial semantics and temporal context clues among agents. Our framework is thoroughly evaluated on several LiDAR-based collaborative detection datasets in real-world and simulated scenarios. Comprehensive experiments demonstrate the superiority of How2comm and the effectiveness of all its vital components. The code will be released at https://github.com/ydk122024/How2comm.

Recently, the transformer has enabled the speed-oriented trackers to approach state-of-the-art (SOTA) performance with high-speed thanks to the smaller input size or the lighter feature extraction backbone, though they still substantially lag behind their corresponding performance-oriented versions. In this paper, we demonstrate that it is possible to narrow or even close this gap while achieving high tracking speed based on the smaller input size. To this end, we non-uniformly resize the cropped image to have a smaller input size while the resolution of the area where the target is more likely to appear is higher and vice versa. This enables us to solve the dilemma of attending to a larger visual field while retaining more raw information for the target despite a smaller input size. Our formulation for the non-uniform resizing can be efficiently solved through quadratic programming (QP) and naturally integrated into most of the crop-based local trackers. Comprehensive experiments on five challenging datasets based on two kinds of transformer trackers, \ie, OSTrack and TransT, demonstrate consistent improvements over them. In particular, applying our method to the speed-oriented version of OSTrack even outperforms its performance-oriented counterpart by 0.6\% AUC on TNL2K, while running 50\% faster and saving over 55\% MACs. Codes and models are available at https://github.com/Kou-99/ZoomTrack.

Recent years have witnessed significant progress in understanding the relationship between the connectivity of a deep network's architecture as a graph, and the network's performance. A few prior arts connected deep architectures to expander graphs or Ramanujan graphs, and particularly,[7] demonstrated the use of such graph connectivity measures with ranking and relative performance of various obtained sparse sub-networks (i.e. models with prune masks) without the need for training. However, no prior work explicitly explores the role of parameters in the graph's connectivity, making the graph-based understanding of prune masks and the magnitude/gradient-based pruning practice isolated from one another. This paper strives to fill in this gap, by analyzing the Weighted Spectral Gap of Ramanujan structures in sparse neural networks and investigates its correlation with final performance. We specifically examine the evolution of sparse structures under a popular dynamic sparse-to-sparse network training scheme, and intriguingly find that the generated random topologies inherently maximize Ramanujan graphs. We also identify a strong correlation between masks, performance, and the weighted spectral gap. Leveraging this observation, we propose to construct a new "full-spectrum coordinate'' aiming to comprehensively characterize a sparse neural network's promise. Concretely, it consists of the classical Ramanujan's gap (structure), our proposed weighted spectral gap (parameters), and the constituent nested regular graphs within. In this new coordinate system, a sparse subnetwork's L2-distance from its original initialization is found to have nearly linear correlated with its performance. Eventually, we apply this unified perspective to develop a new actionable pruning method, by sampling sparse masks to maximize the L2-coordinate distance. Our method can be augmented with the "pruning at initialization" (PaI) method, and significantly outperforms existing PaI methods. With only a few iterations of training (e.g 500 iterations), we can get LTH-comparable performance as that yielded via "pruning after training", significantly saving pre-training costs. Codes can be found at: https://github.com/VITA-Group/FullSpectrum-PAI.

In privacy under continual observation we study how to release differentially private estimates based on a dataset that evolves over time. The problem of releasing private prefix sums of $x_1, x_2, x_3,\dots\in${$0,1$} (where the value of each $x_i$ is to be private) is particularly well-studied, and a generalized form is used in state-of-the-art methods for private stochastic gradient descent (SGD). The seminal binary mechanism privately releases the first $t$ prefix sums with noise of variance polylogarithmic in $t$. Recently, Henzinger et al. and Denisov et al. showed that it is possible to improve on the binary mechanism in two ways: The variance of the noise can be reduced by a (large) constant factor, and also made more even across time steps. However, their algorithms for generating the noise distribution are not as efficient as one would like in terms of computation time and (in particular) space. We address the efficiency problem by presenting a simple alternative to the binary mechanism in which 1) generating the noise takes constant average time per value, 2) the variance is reduced by a factor about 4 compared to the binary mechanism, and 3) the noise distribution at each step is identical. Empirically, a simple Python implementation of our approach outperforms the running time of the approach of Henzinger et al., as well as an attempt to improve their algorithm using high-performance algorithms for multiplication with Toeplitz matrices.

Existing dataset distillation (DD) techniques typically rely on iterative strategies to synthesize condensed datasets, where datasets before and after distillation are forward and backward through neural networks a massive number of times. Despite the promising results achieved, the time efficiency of prior approaches is still far from satisfactory. Moreover, when different sizes of synthetic datasets are required, they have to repeat the iterative training procedures, which is highly cumbersome and lacks flexibility. In this paper, different from the time-consuming forward-backward passes, we introduce a generative fashion for dataset distillation with significantly improved efficiency. Specifically, synthetic samples are produced by a generator network conditioned on the initialization of DD, while synthetic labels are obtained by solving a least-squares problem in a feature space. Our theoretical analysis reveals that the errors of synthetic datasets solved in the original space and then processed by any conditional generators are upper-bounded. To find a satisfactory generator efficiently, we propose a meta-learning algorithm, where a meta generator is trained on a large dataset so that only a few steps are required to adapt to a target dataset. The meta generator is termed as MGDD in our approach. Once adapted, it can handle arbitrary sizes of synthetic datasets, even for those unseen during adaptation. Experiments demonstrate that the generator adapted with only a limited number of steps performs on par with those state-of-the-art DD methods and yields $22\times$ acceleration.

We propose Neural Priming, a technique for adapting large pretrained models to distribution shifts and downstream tasks given few or no labeled examples. Presented with class names or unlabeled test samples, Neural Priming enables the model to recall and conditions its parameters on relevant data seen throughout pretraining, thereby priming it for the test distribution. Neural Priming can be performed at test time in even for pretraining datasets as large as LAION-2B. Performing lightweight updates on the recalled data significantly improves accuracy across a variety of distribution shift and transfer learning benchmarks. Concretely, in the zero-shot setting, we see a 2.45% improvement in accuracy on ImageNet and 3.81% accuracy improvement on average across standard transfer learning benchmarks. Further, using our test time inference scheme, we see a 1.41% accuracy improvement on ImageNetV2. These results demonstrate the effectiveness of Neural Priming in addressing the common challenge of limited labeled data and changing distributions. Code and models are open-sourced at [https://www.github.com/RAIVNLab/neural-priming](https://www.github.com/RAIVNLab/neural-priming).

Embedding tables are used by machine learning systems to work with categorical features. In modern Recommendation Systems, these tables can be very large, necessitating the development of new methods for fitting them in memory, even during training. We suggest Clustered Compositional Embeddings (CCE) which combines clustering-based compression like quantization to codebooks with dynamic methods like The Hashing Trick and Compositional Embeddings [Shi et al., 2020]. Experimentally CCE achieves the best of both worlds: The high compression rate of codebook-based quantization, but \emph{dynamically} like hashing-based methods, so it can be used during training. Theoretically, we prove that CCE is guaranteed to converge to the optimal codebook and give a tight bound for the number of iterations required.

Recent research has highlighted the importance of dataset size in scaling language models. However, large language models (LLMs) are notoriously token-hungry during pre-training, and high-quality text data on the web is likely to be approaching its scaling limit for LLMs. To further enhance LLMs, a straightforward approach is to repeat the pre-training data for additional epochs. In this study, we empirically investigate three key aspects under this approach. First, we explore the consequences of repeating pre-training data, revealing that the model is susceptible to overfitting, leading to multi-epoch degradation. Second, we examine the key factors contributing to multi-epoch degradation, finding that significant factors include dataset size, model parameters, and training objectives, while less influential factors consist of dataset quality and model FLOPs. Finally, we explore whether widely used regularization can alleviate multi-epoch degradation. Most regularization techniques do not yield significant improvements, except for dropout, which demonstrates remarkable effectiveness but requires careful tuning when scaling up the model size. Additionally, we discover that leveraging mixture-of-experts (MoE) enables cost-effective and efficient hyper-parameter tuning for computationally intensive dense LLMs with comparable trainable parameters, potentially impacting efficient LLM development on a broader scale.

LiDAR-based 3D detection methods currently use bird's-eye view (BEV) or range view (RV) as their primary basis. The former relies on voxelization and 3D convolutions, resulting in inefficient training and inference processes. Conversely, RV-based methods demonstrate higher efficiency due to their compactness and compatibility with 2D convolutions, but their performance still trails behind that of BEV-based methods. To eliminate this performance gap while preserving the efficiency of RV-based methods, this study presents an efficient and accurate RV-based 3D object detection framework termed RangePerception. Through meticulous analysis, this study identifies two critical challenges impeding the performance of existing RV-based methods: 1) there exists a natural domain gap between the 3D world coordinate used in output and 2D range image coordinate used in input, generating difficulty in information extraction from range images; 2) native range images suffer from vision corruption issue, affecting the detection accuracy of the objects located on the margins of the range images. To address the key challenges above, we propose two novel algorithms named Range Aware Kernel (RAK) and Vision Restoration Module (VRM), which facilitate information flow from range image representation and world-coordinate 3D detection results. With the help of RAK and VRM, our RangePerception achieves 3.25/4.18 higher averaged L1/L2 AP compared to previous state-of-the-art RV-based method RangeDet, on Waymo Open Dataset. For the first time as an RV-based 3D detection method, RangePerception achieves slightly superior averaged AP compared with the well-known BEV-based method CenterPoint and the inference speed of RangePerception is 1.3 times as fast as CenterPoint.

Transformer-based trackers have achieved strong accuracy on the standard benchmarks. However, their efficiency remains an obstacle to practical deployment on both GPU and CPU platforms. In this paper, to overcome this issue, we propose a fully transformer tracking framework, coined as \emph{MixFormerV2}, without any dense convolutional operation and complex score prediction module. Our key design is to introduce four special prediction tokens and concatenate them with the tokens from target template and search areas. Then, we apply the unified transformer backbone on these mixed token sequence. These prediction tokens are able to capture the complex correlation between target template and search area via mixed attentions. Based on them, we can easily predict the tracking box and estimate its confidence score through simple MLP heads. To further improve the efficiency of MixFormerV2, we present a new distillation-based model reduction paradigm, including dense-to-sparse distillation and deep-to-shallow distillation. The former one aims to transfer knowledge from the dense-head based MixViT to our fully transformer tracker, while the latter one is used to prune some layers of the backbone. We instantiate two types of MixForemrV2, where the MixFormerV2-B achieves an AUC of 70.6\% on LaSOT and AUC of 56.7\% on TNL2k with a high GPU speed of 165 FPS, and the MixFormerV2-S surpasses FEAR-L by 2.7\% AUC on LaSOT with a real-time CPU speed.

Spiking Neural Networks (SNNs) are considered promising brain-inspired energy-efficient models due to their event-driven computing paradigm. The spatiotemporal spike patterns used to convey information in SNNs consist of both rate coding and temporal coding, where the temporal coding is crucial to biological-plausible learning rules such as spike-timing-dependent-plasticity. The time-based training strategy is proposed to better utilize the temporal information in SNNs and learn in an asynchronous fashion. However, some recent works train SNNs by the time-based scheme with rate-coding-dominated loss functions. In this paper, we first map rate-based loss functions to time-based counterparts and explain why they are also applicable to the time-based training scheme. After that, we infer that loss functions providing adequate positive overall gradients help training by theoretical analysis. Based on this, we propose the enhanced counting loss to replace the commonly used mean square counting loss. In addition, we transfer the training of scale factor in weight standardization into thresholds. Experiments show that our approach outperforms previous time-based training methods in most datasets. Our work provides insights for training SNNs with time-based schemes and offers a fresh perspective on the correlation between rate coding and temporal coding. Our code is available at https://github.com/zhuyaoyu/SNN-temporal-training-losses.

Subgame solving is an essential technique in addressing large imperfect information games, with various approaches developed to enhance the performance of refined strategies in the abstraction of the target subgame. However, directly applying existing subgame solving techniques may be difficult, due to the intricate nature and substantial size of many real-world games. To overcome this issue, recent subgame solving methods allow for subgame solving on limited knowledge order subgames, increasing their applicability in large games; yet this may still face obstacles due to extensive information set sizes. To address this challenge, we propose a generative subgame solving (GS2) framework, which utilizes a generation function to identify a subset of the earliest-reached nodes, reducing the size of the subgame. Our method is supported by a theoretical analysis and employs a diversity-based generation function to enhance safety. Experiments conducted on medium-sized games as well as the challenging large game of GuanDan demonstrate a significant improvement over the blueprint.

An important goal of modern scheduling systems is to efficiently manage power usage. In energy-efficient scheduling, the operating system controls the speed at which a machine is processing jobs with the dual objective of minimizing energy consumption and optimizing the quality of service cost of the resulting schedule. Since machine-learned predictions about future requests can often be learned from historical data, a recent line of work on learning-augmented algorithms aims to achieve improved performance guarantees by leveraging predictions. In particular, for energy-efficient scheduling, Bamas et. al. [NeurIPS '20] and Antoniadis et. al. [SWAT '22] designed algorithms with predictions for the energy minimization with deadlines problem and achieved an improved competitive ratio when the prediction error is small while also maintaining worst-case bounds even when the prediction error is arbitrarily large. In this paper, we consider a general setting for energy-efficient scheduling and provide a flexible learning-augmented algorithmic framework that takes as input an offline and an online algorithm for the desired energy-efficient scheduling problem. We show that, when the prediction error is small, this framework gives improved competitive ratios for many different energy-efficient scheduling problems, including energy minimization with deadlines, while also maintaining a bounded competitive ratio regardless of the prediction error. Finally, we empirically demonstrate that this framework achieves an improved performance on real and synthetic datasets.

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.

Since developing a new multimodal LLM (MLLM) by pre-training on tremendous image-text pairs from scratch can be exceedingly resource-consuming, connecting an existing LLM with a comparatively lightweight visual prompt generator (VPG) becomes a feasible paradigm. However, further tuning the VPG component of the MLLM still incurs significant computational costs, such as thousands of GPU hours and millions of training data points. An alternative solution is transferring an existing VPG from one MLLM to the target MLLM. In this work, we investigate VPG transferability across LLMs for the first time, aiming to reduce the cost of VPG training. Specifically, we explore VPG transfer across different LLM sizes (e.g., small-to-large) and types. We identify key factors to maximize transfer efficiency, based on which we develop a simple yet highly effective two-stage transfer framework, called VPGTrans. Notably, it enables VPG transfer from BLIP-2 OPT 2.7B to BLIP-2 OPT 6.7B with less than 10% of the GPU hours using only 10.7% of the training data compared to training a VPG for OPT 6.7B from scratch. Furthermore, we provide a series of intriguing findings and discuss potential explanations behind them. Finally, we showcase the practical value of our VPGTrans approach, by customizing two novel MLLMs, including VL-LLaMA and VL-Vicuna, with recently released LLaMA and Vicuna LLMs.

Video Semantic Segmentation (VSS) involves assigning a semantic label to each pixel in a video sequence. Prior work in this field has demonstrated promising results by extending image semantic segmentation models to exploit temporal relationships across video frames; however, these approaches often incur significant computational costs. In this paper, we propose an efficient mask propagation framework for VSS, called MPVSS. Our approach first employs a strong query-based image segmentor on sparse key frames to generate accurate binary masks and class predictions. We then design a flow estimation module utilizing the learned queries to generate a set of segment-aware flow maps, each associated with a mask prediction from the key frame. Finally, the mask-flow pairs are warped to serve as the mask predictions for the non-key frames. By reusing predictions from key frames, we circumvent the need to process a large volume of video frames individually with resource-intensive segmentors, alleviating temporal redundancy and significantly reducing computational costs. Extensive experiments on VSPW and Cityscapes demonstrate that our mask propagation framework achieves SOTA accuracy and efficiency trade-offs. For instance, our best model with Swin-L backbone outperforms the SOTA MRCFA using MiT-B5 by 4.0% mIoU, requiring only 26% FLOPs on the VSPW dataset. Moreover, our framework reduces up to 4× FLOPs compared to the per-frame Mask2Former baseline with only up to 2% mIoU degradation on the Cityscapes validation set. Code is available at https://github.com/ziplab/MPVSS.