This paper explores the generalization characteristics of iterative learning algorithms with bounded updates for non-convex loss functions, employing information-theoretic techniques. Our key contribution is a novel bound for the generalization error of these algorithms with bounded updates, extending beyond the scope of previous works that only focused on Stochastic Gradient Descent (SGD). Our approach introduces two main novelties: 1) we reformulate the mutual information as the uncertainty of updates, providing a new perspective, and 2) instead of using the chaining rule of mutual information, we employ a variance decomposition technique to decompose information across iterations, allowing for a simpler surrogate process. We analyze our generalization bound under various settings and demonstrate improved bounds when the model dimension increases at the same rate as the number of training data samples. To bridge the gap between theory and practice, we also examine the previously observed scaling behavior in large language models. Ultimately, our work takes a further step for developing practical generalization theories.
In-context learning, i.e., learning from in-context samples, is an impressive ability of Transformer. However, the mechanism driving the in-context learning is not yet fully understood. In this study, we aim to investigate from an underexplored perspective of representation learning. The representation is more complex for in-context learning senario, where the representation can be impacted by both model weights and in-context samples. We refer the above two conceptually aspects of representation as in-weight component and in-context component, respectively. To study how the two components affect in-context learning capabilities, we construct a novel synthetic task, making it possible to device two probes, in-weights probe and in-context probe, to evaluate the two components, respectively. We demonstrate that the goodness of in-context component is highly related to the in-context learning performance, which indicates the entanglement between in-context learning and representation learning. Furthermore, we find that a good in-weights component can actually benefit the learning of the in-context component, indicating that in-weights learning should be the foundation of in-context learning. To further understand the the in-context learning mechanism and importance of the in-weights component, we proof by construction that a simple Transformer, which uses pattern matching and copy-past mechanism to perform in-context learning, can match the in-context learning performance with more complex, best tuned Transformer under the perfect in-weights component assumption. In short, those discoveries from representation learning perspective shed light on new approaches to improve the in-context capacity.
Planning and prediction are two important modules of autonomous driving and have experienced tremendous advancement recently. Nevertheless, most existing methods regard planning and prediction as independent and ignore the correlation between them, leading to the lack of consideration for interaction and dynamic changes of traffic scenarios. To address this challenge, we propose InteractionNet, which leverages transformer to share global contextual reasoning among all traffic participants to capture interaction and interconnect planning and prediction to achieve joint. Besides, InteractionNet deploys another transformer to help the model pay extra attention to the perceived region containing critical or unseen vehicles. InteractionNet outperforms other baselines in several benchmarks, especially in terms of safety, which benefits from the joint consideration of planning and forecasting. The code will be available at https://github.com/fujiawei0724/InteractionNet.
High-Definition (HD) maps play a crucial role in autonomous driving systems. Recent methods have attempted to construct HD maps in real-time based on information obtained from vehicle onboard sensors. However, the performance of these methods is significantly susceptible to the environment surrounding the vehicle due to the inherent limitation of onboard sensors, such as weak capacity for long-range detection. In this study, we demonstrate that supplementing onboard sensors with satellite maps can enhance the performance of HD map construction methods, leveraging the broad coverage capability of satellite maps. For the purpose of further research, we release the satellite map tiles as a complementary dataset of nuScenes dataset. Meanwhile, we propose a hierarchical fusion module that enables better fusion of satellite maps information with existing methods. Specifically, we design an attention mask based on segmentation and distance, applying the cross-attention mechanism to fuse onboard Bird's Eye View (BEV) features and satellite features in feature-level fusion. An alignment module is introduced before concatenation in BEV-level fusion to mitigate the impact of misalignment between the two features. The experimental results on the augmented nuScenes dataset showcase the seamless integration of our module into three existing HD map construction methods. It notably enhances their performance in both HD map semantic segmentation and instance detection tasks.
We introduce a highly performant 3D object detector for point clouds using the DETR framework. The prior attempts all end up with suboptimal results because they fail to learn accurate inductive biases from the limited scale of training data. In particular, the queries often attend to points that are far away from the target objects, violating the locality principle in object detection. To address the limitation, we introduce a novel 3D Vertex Relative Position Encoding (3DV-RPE) method which computes position encoding for each point based on its relative position to the 3D boxes predicted by the queries in each decoder layer, thus providing clear information to guide the model to focus on points near the objects, in accordance with the principle of locality. In addition, we systematically improve the pipeline from various aspects such as data normalization based on our understanding of the task. We show exceptional results on the challenging ScanNetV2 benchmark, achieving significant improvements over the previous 3DETR in $\rm{AP}_{25}$/$\rm{AP}_{50}$ from 65.0\%/47.0\% to 77.8\%/66.0\%, respectively. In addition, our method sets a new record on ScanNetV2 and SUN RGB-D datasets.Code will be released at http://github.com/yichaoshen-MS/V-DETR.
This paper presents an improved DETR detector that maintains a "plain" nature: using a single-scale feature map and global cross-attention calculations without specific locality constraints, in contrast to previous leading DETR-based detectors that reintroduce architectural inductive biases of multi-scale and locality into the decoder. We show that two simple technologies are surprisingly effective within a plain design to compensate for the lack of multi-scale feature maps and locality constraints. The first is a box-to-pixel relative position bias (BoxRPB) term added to the cross-attention formulation, which well guides each query to attend to the corresponding object region while also providing encoding flexibility. The second is masked image modeling (MIM)-based backbone pre-training which helps learn representation with fine-grained localization ability and proves crucial for remedying dependencies on the multi-scale feature maps. By incorporating these technologies and recent advancements in training and problem formation, the improved "plain" DETR showed exceptional improvements over the original DETR detector. By leveraging the Object365 dataset for pre-training, it achieved 63.9 mAP accuracy using a Swin-L backbone, which is highly competitive with state-of-the-art detectors which all heavily rely on multi-scale feature maps and region-based feature extraction. Code is available at https://github.com/impiga/Plain-DETR .
It has long been an ill-posed problem to predict absolute depth maps from single images in real (unseen) indoor scenes. We observe that it is essentially due to not only the scale-ambiguous problem but also the focal-ambiguous problem that decreases the generalization ability of monocular depth estimation. That is, images may be captured by cameras of different focal lengths in scenes of different scales. In this paper, we develop a focal-and-scale depth estimation model to well learn absolute depth maps from single images in unseen indoor scenes. First, a relative depth estimation network is adopted to learn relative depths from single images with diverse scales/semantics. Second, multi-scale features are generated by mapping a single focal length value to focal length features and concatenating them with intermediate features of different scales in relative depth estimation. Finally, relative depths and multi-scale features are jointly fed into an absolute depth estimation network. In addition, a new pipeline is developed to augment the diversity of focal lengths of public datasets, which are often captured with cameras of the same or similar focal lengths. Our model is trained on augmented NYUDv2 and tested on three unseen datasets. Our model considerably improves the generalization ability of depth estimation by 41%/13% (RMSE) with/without data augmentation compared with five recent SOTAs and well alleviates the deformation problem in 3D reconstruction. Notably, our model well maintains the accuracy of depth estimation on original NYUDv2.
Scaling sequence length has become a critical demand in the era of large language models. However, existing methods struggle with either computational complexity or model expressivity, rendering the maximum sequence length restricted. To address this issue, we introduce LongNet, a Transformer variant that can scale sequence length to more than 1 billion tokens, without sacrificing the performance on shorter sequences. Specifically, we propose dilated attention, which expands the attentive field exponentially as the distance grows. LongNet has significant advantages: 1) it has a linear computation complexity and a logarithm dependency between any two tokens in a sequence; 2) it can be served as a distributed trainer for extremely long sequences; 3) its dilated attention is a drop-in replacement for standard attention, which can be seamlessly integrated with the existing Transformer-based optimization. Experiments results demonstrate that LongNet yields strong performance on both long-sequence modeling and general language tasks. Our work opens up new possibilities for modeling very long sequences, e.g., treating a whole corpus or even the entire Internet as a sequence.
In this paper, we propose a novel and effective Multi-Level Fusion network, named as MLF-DET, for high-performance cross-modal 3D object DETection, which integrates both the feature-level fusion and decision-level fusion to fully utilize the information in the image. For the feature-level fusion, we present the Multi-scale Voxel Image fusion (MVI) module, which densely aligns multi-scale voxel features with image features. For the decision-level fusion, we propose the lightweight Feature-cued Confidence Rectification (FCR) module which further exploits image semantics to rectify the confidence of detection candidates. Besides, we design an effective data augmentation strategy termed Occlusion-aware GT Sampling (OGS) to reserve more sampled objects in the training scenes, so as to reduce overfitting. Extensive experiments on the KITTI dataset demonstrate the effectiveness of our method. Notably, on the extremely competitive KITTI car 3D object detection benchmark, our method reaches 82.89% moderate AP and achieves state-of-the-art performance without bells and whistles.
Momentum has become a crucial component in deep learning optimizers, necessitating a comprehensive understanding of when and why it accelerates stochastic gradient descent (SGD). To address the question of ''when'', we establish a meaningful comparison framework that examines the performance of SGD with Momentum (SGDM) under the \emph{effective learning rates} $\eta_{ef}$, a notion unifying the influence of momentum coefficient $\mu$ and batch size $b$ over learning rate $\eta$. In the comparison of SGDM and SGD with the same effective learning rate and the same batch size, we observe a consistent pattern: when $\eta_{ef}$ is small, SGDM and SGD experience almost the same empirical training losses; when $\eta_{ef}$ surpasses a certain threshold, SGDM begins to perform better than SGD. Furthermore, we observe that the advantage of SGDM over SGD becomes more pronounced with a larger batch size. For the question of ``why'', we find that the momentum acceleration is closely related to \emph{abrupt sharpening} which is to describe a sudden jump of the directional Hessian along the update direction. Specifically, the misalignment between SGD and SGDM happens at the same moment that SGD experiences abrupt sharpening and converges slower. Momentum improves the performance of SGDM by preventing or deferring the occurrence of abrupt sharpening. Together, this study unveils the interplay between momentum, learning rates, and batch sizes, thus improving our understanding of momentum acceleration.