Mixture of Experts (MoE) is able to scale up vision transformers effectively. However, it requires prohibiting computation resources to train a large MoE transformer. In this paper, we propose Residual Mixture of Experts (RMoE), an efficient training pipeline for MoE vision transformers on downstream tasks, such as segmentation and detection. RMoE achieves comparable results with the upper-bound MoE training, while only introducing minor additional training cost than the lower-bound non-MoE training pipelines. The efficiency is supported by our key observation: the weights of an MoE transformer can be factored into an input-independent core and an input-dependent residual. Compared with the weight core, the weight residual can be efficiently trained with much less computation resource, e.g., finetuning on the downstream data. We show that, compared with the current MoE training pipeline, we get comparable results while saving over 30% training cost. When compared with state-of-the-art non- MoE transformers, such as Swin-T / CvT-13 / Swin-L, we get +1.1 / 0.9 / 1.0 mIoU gain on ADE20K segmentation and +1.4 / 1.6 / 0.6 AP gain on MS-COCO object detection task with less than 3% additional training cost.
相關內容
Artificial neural network (ANN) is a versatile tool to study the neural representation in the ventral visual stream, and the knowledge in neuroscience in return inspires ANN models to improve performance in the task. However, it is still unclear how to merge these two directions into a unified framework. In this study, we propose an integrated framework called Deep Autoencoder with Neural Response (DAE-NR), which incorporates information from ANN and the visual cortex to achieve better image reconstruction performance and higher neural representation similarity between biological and artificial neurons. The same visual stimuli (i.e., natural images) are input to both the mice brain and DAE-NR. The encoder of DAE-NR jointly learns the dependencies from neural spike encoding and image reconstruction. For the neural spike encoding task, the features derived from a specific hidden layer of the encoder are transformed by a mapping function to predict the ground-truth neural response under the constraint of image reconstruction. Simultaneously, for the image reconstruction task, the latent representation obtained by the encoder is assigned to a decoder to restore the original image under the guidance of neural information. In DAE-NR, the learning process of encoder, mapping function and decoder are all implicitly constrained by these two tasks. Our experiments demonstrate that if and only if with the joint learning, DAE-NRs can improve the performance of visual image reconstruction and increase the representation similarity between biological neurons and artificial neurons. The DAE-NR offers a new perspective on the integration of computer vision and neuroscience.
To support the application scenarios where high-resolution (HR) images are urgently needed, various single image super-resolution (SISR) algorithms are developed. However, SISR is an ill-posed inverse problem, which may bring artifacts like texture shift, blur, etc. to the reconstructed images, thus it is necessary to evaluate the quality of super-resolution images (SRIs). Note that most existing image quality assessment (IQA) methods were developed for synthetically distorted images, which may not work for SRIs since their distortions are more diverse and complicated. Therefore, in this paper, we propose a no-reference deep-learning image quality assessment method based on frequency maps because the artifacts caused by SISR algorithms are quite sensitive to frequency information. Specifically, we first obtain the high-frequency map (HM) and low-frequency map (LM) of SRI by using Sobel operator and piecewise smooth image approximation. Then, a two-stream network is employed to extract the quality-aware features of both frequency maps. Finally, the features are regressed into a single quality value using fully connected layers. The experimental results show that our method outperforms all compared IQA models on the selected three super-resolution quality assessment (SRQA) databases.
Domain generalization (DG) aims at learning generalizable models under distribution shifts to avoid redundantly overfitting massive training data. Previous works with complex loss design and gradient constraint have not yet led to empirical success on large-scale benchmarks. In this work, we reveal the mixture-of-experts (MoE) model's generalizability on DG by leveraging to distributively handle multiple aspects of the predictive features across domains. To this end, we propose Sparse Fusion Mixture-of-Experts (SF-MoE), which incorporates sparsity and fusion mechanisms into the MoE framework to keep the model both sparse and predictive. SF-MoE has two dedicated modules: 1) sparse block and 2) fusion block, which disentangle and aggregate the diverse learned signals of an object, respectively. Extensive experiments demonstrate that SF-MoE is a domain-generalizable learner on large-scale benchmarks. It outperforms state-of-the-art counterparts by more than 2% across 5 large-scale DG datasets (e.g., DomainNet), with the same or even lower computational costs. We further reveal the internal mechanism of SF-MoE from distributed representation perspective (e.g., visual attributes). We hope this framework could facilitate future research to push generalizable object recognition to the real world. Code and models are released at //github.com/Luodian/SF-MoE-DG.
In experiments that study social phenomena, such as peer influence or herd immunity, the treatment of one unit may influence the outcomes of others. Such "interference between units" violates traditional approaches for causal inference, so that additional assumptions are often imposed to model or limit the underlying social mechanism. For binary outcomes, we propose an approach that does not require such assumptions, allowing for interference that is both unmodeled and strong, with confidence intervals derived using only the randomization of treatment. However, the estimates will have wider confidence intervals and weaker causal implications than those attainable under stronger assumptions. The approach allows for the usage of regression, matching, or weighting, as may best fit the application at hand. Inference is done by bounding the distribution of the estimation error over all possible values of the unknown counterfactual, using an integer program. Examples are shown using using a vaccination trial and two experiments investigating social influence.
In recent years, Mixture-of-Experts (MoE) has emerged as a promising technique for deep learning that can scale the model capacity to trillion-plus parameters while reducing the computing cost via sparse computation. While MoE opens a new frontier of exceedingly large models, its implementation over thousands of GPUs has been limited due to mismatch between the dynamic nature of MoE and static parallelism/pipelining of the system. We present Tutel, a highly scalable stack design and implementation for MoE with dynamically adaptive parallelism and pipelining. Tutel delivers adaptive parallelism switching and adaptive pipelining at runtime, which achieves up to 1.74x and 2.00x single MoE layer speedup, respectively. We also propose a novel two-dimensional hierarchical algorithm for MoE communication speedup that outperforms the previous state-of-the-art up to 20.7x over 2,048 GPUs. Aggregating all techniques, Tutel finally delivers 4.96x and 5.75x speedup of a single MoE layer on 16 GPUs and 2,048 GPUs, respectively, over Fairseq: Meta's Facebook AI Research Sequence-to-Sequence Toolkit (Tutel is now partially adopted by Fairseq). Tutel source code is available in public: //github.com/microsoft/tutel . Our evaluation shows that Tutel efficiently and effectively runs a real-world MoE-based model named SwinV2-MoE, built upon Swin Transformer V2, a state-of-the-art computer vision architecture. On efficiency, Tutel accelerates SwinV2-MoE, achieving up to 1.55x and 2.11x speedup in training and inference over Fairseq, respectively. On effectiveness, the SwinV2-MoE model achieves superior accuracy in both pre-training and down-stream computer vision tasks such as COCO object detection than the counterpart dense model, indicating the readiness of Tutel for end-to-end real-world model training and inference. SwinV2-MoE is open sourced in //github.com/microsoft/Swin-Transformer .
Predicting fetal weight at birth is an important aspect of perinatal care, particularly in the context of antenatal management, which includes the planned timing and the mode of delivery. Accurate prediction of weight using prenatal ultrasound is challenging as it requires images of specific fetal body parts during advanced pregnancy which is difficult to capture due to poor quality of images caused by the lack of amniotic fluid. As a consequence, predictions which rely on standard methods often suffer from significant errors. In this paper we propose the Residual Transformer Module which extends a 3D ResNet-based network for analysis of 2D+t spatio-temporal ultrasound video scans. Our end-to-end method, called BabyNet, automatically predicts fetal birth weight based on fetal ultrasound video scans. We evaluate BabyNet using a dedicated clinical set comprising 225 2D fetal ultrasound videos of pregnancies from 75 patients performed one day prior to delivery. Experimental results show that BabyNet outperforms several state-of-the-art methods and estimates the weight at birth with accuracy comparable to human experts. Furthermore, combining estimates provided by human experts with those computed by BabyNet yields the best results, outperforming either of other methods by a significant margin. The source code of BabyNet is available at //github.com/SanoScience/BabyNet.
Residual networks (ResNets) have displayed impressive results in pattern recognition and, recently, have garnered considerable theoretical interest due to a perceived link with neural ordinary differential equations (neural ODEs). This link relies on the convergence of network weights to a smooth function as the number of layers increases. We investigate the properties of weights trained by stochastic gradient descent and their scaling with network depth through detailed numerical experiments. We observe the existence of scaling regimes markedly different from those assumed in neural ODE literature. Depending on certain features of the network architecture, such as the smoothness of the activation function, one may obtain an alternative ODE limit, a stochastic differential equation or neither of these. These findings cast doubts on the validity of the neural ODE model as an adequate asymptotic description of deep ResNets and point to an alternative class of differential equations as a better description of the deep network limit.
In multi-turn dialog, utterances do not always take the full form of sentences \cite{Carbonell1983DiscoursePA}, which naturally makes understanding the dialog context more difficult. However, it is essential to fully grasp the dialog context to generate a reasonable response. Hence, in this paper, we propose to improve the response generation performance by examining the model's ability to answer a reading comprehension question, where the question is focused on the omitted information in the dialog. Enlightened by the multi-task learning scheme, we propose a joint framework that unifies these two tasks, sharing the same encoder to extract the common and task-invariant features with different decoders to learn task-specific features. To better fusing information from the question and the dialog history in the encoding part, we propose to augment the Transformer architecture with a memory updater, which is designed to selectively store and update the history dialog information so as to support downstream tasks. For the experiment, we employ human annotators to write and examine a large-scale dialog reading comprehension dataset. Extensive experiments are conducted on this dataset, and the results show that the proposed model brings substantial improvements over several strong baselines on both tasks. In this way, we demonstrate that reasoning can indeed help better response generation and vice versa. We release our large-scale dataset for further research.
Deep convolutional neural networks (CNNs) have recently achieved great success in many visual recognition tasks. However, existing deep neural network models are computationally expensive and memory intensive, hindering their deployment in devices with low memory resources or in applications with strict latency requirements. Therefore, a natural thought is to perform model compression and acceleration in deep networks without significantly decreasing the model performance. During the past few years, tremendous progress has been made in this area. In this paper, we survey the recent advanced techniques for compacting and accelerating CNNs model developed. These techniques are roughly categorized into four schemes: parameter pruning and sharing, low-rank factorization, transferred/compact convolutional filters, and knowledge distillation. Methods of parameter pruning and sharing will be described at the beginning, after that the other techniques will be introduced. For each scheme, we provide insightful analysis regarding the performance, related applications, advantages, and drawbacks etc. Then we will go through a few very recent additional successful methods, for example, dynamic capacity networks and stochastic depths networks. After that, we survey the evaluation matrix, the main datasets used for evaluating the model performance and recent benchmarking efforts. Finally, we conclude this paper, discuss remaining challenges and possible directions on this topic.
This paper introduces an online model for object detection in videos designed to run in real-time on low-powered mobile and embedded devices. Our approach combines fast single-image object detection with convolutional long short term memory (LSTM) layers to create an interweaved recurrent-convolutional architecture. Additionally, we propose an efficient Bottleneck-LSTM layer that significantly reduces computational cost compared to regular LSTMs. Our network achieves temporal awareness by using Bottleneck-LSTMs to refine and propagate feature maps across frames. This approach is substantially faster than existing detection methods in video, outperforming the fastest single-frame models in model size and computational cost while attaining accuracy comparable to much more expensive single-frame models on the Imagenet VID 2015 dataset. Our model reaches a real-time inference speed of up to 15 FPS on a mobile CPU.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191