SE-MoE: A Scalable and Efficient Mixture-of-Experts Distributed Training and Inference System

With the increasing diversity of ML infrastructures nowadays, distributed training over heterogeneous computing systems is desired to facilitate the production of big models. Mixture-of-Experts (MoE) models have been proposed to lower the cost of training subject to the overall size of models/data through gating and parallelism in a divide-and-conquer fashion. While DeepSpeed has made efforts in carrying out large-scale MoE training over heterogeneous infrastructures, the efficiency of training and inference could be further improved from several system aspects, including load balancing, communication/computation efficiency, and memory footprint limits. In this work, we present SE-MoE that proposes Elastic MoE training with 2D prefetch and Fusion communication over Hierarchical storage, so as to enjoy efficient parallelisms in various types. For scalable inference in a single node, especially when the model size is larger than GPU memory, SE-MoE forms the CPU-GPU memory jointly into a ring of sections to load the model, and executes the computation tasks across the memory sections in a round-robin manner for efficient inference. We carried out extensive experiments to evaluate SE-MoE, where SE-MoE successfully trains a Unified Feature Optimization (UFO) model with a Sparsely-Gated Mixture-of-Experts model of 12B parameters in 8 days on 48 A100 GPU cards. The comparison against the state-of-the-art shows that SE-MoE outperformed DeepSpeed with 33% higher throughput (tokens per second) in training and 13% higher throughput in inference in general. Particularly, under unbalanced MoE Tasks, e.g., UFO, SE-MoE achieved 64% higher throughput with 18% lower memory footprints. The code of the framework will be released on: https://github.com/PaddlePaddle/Paddle.

A Comparative Survey of Deep Active Learning

Active Learning (AL) is a set of techniques for reducing labeling cost by sequentially selecting data samples from a large unlabeled data pool for labeling. Meanwhile, Deep Learning (DL) is data-hungry, and the performance of DL models scales monotonically with more training data. Therefore, in recent years, Deep Active Learning (DAL) has risen as feasible solutions for maximizing model performance while minimizing the expensive labeling cost. Abundant methods have sprung up and literature reviews of DAL have been presented before. However, the performance comparison of different branches of DAL methods under various tasks is still insufficient and our work fills this gap. In this paper, we survey and categorize DAL-related work and construct comparative experiments across frequently used datasets and DAL algorithms. Additionally, we explore some factors (e.g., batch size, number of epochs in the training process) that influence the efficacy of DAL, which provides better references for researchers to design their own DAL experiments or carry out DAL-related applications. We construct a DAL toolkit, DeepAL+, by re-implementing many highly-cited DAL-related methods, and it will be released to the public.

PP-HumanSeg: Connectivity-Aware Portrait Segmentation with a Large-Scale Teleconferencing Video Dataset

As the COVID-19 pandemic rampages across the world, the demands of video conferencing surge. To this end, real-time portrait segmentation becomes a popular feature to replace backgrounds of conferencing participants. While feature-rich datasets, models and algorithms have been offered for segmentation that extract body postures from life scenes, portrait segmentation has yet not been well covered in a video conferencing context. To facilitate the progress in this field, we introduce an open-source solution named PP-HumanSeg. This work is the first to construct a large-scale video portrait dataset that contains 291 videos from 23 conference scenes with 14K fine-labeled frames and extensions to multi-camera teleconferencing. Furthermore, we propose a novel Semantic Connectivity-aware Learning (SCL) for semantic segmentation, which introduces a semantic connectivity-aware loss to improve the quality of segmentation results from the perspective of connectivity. And we propose an ultra-lightweight model with SCL for practical portrait segmentation, which achieves the best trade-off between IoU and the speed of inference. Extensive evaluations on our dataset demonstrate the superiority of SCL and our model. The source code is available at https://github.com/PaddlePaddle/PaddleSeg.

SenseMag: Enabling Low-Cost Traffic Monitoring using Non-invasive Magnetic Sensing

The operation and management of intelligent transportation systems (ITS), such as traffic monitoring, relies on real-time data aggregation of vehicular traffic information, including vehicular types (e.g., cars, trucks, and buses), in the critical roads and highways. While traditional approaches based on vehicular-embedded GPS sensors or camera networks would either invade drivers' privacy or require high deployment cost, this paper introduces a low-cost method, namely SenseMag, to recognize the vehicular type using a pair of non-invasive magnetic sensors deployed on the straight road section. SenseMag filters out noises and segments received magnetic signals by the exact time points that the vehicle arrives or departs from every sensor node. Further, SenseMag adopts a hierarchical recognition model to first estimate the speed/velocity, then identify the length of vehicle using the predicted speed, sampling cycles, and the distance between the sensor nodes. With the vehicle length identified and the temporal/spectral features extracted from the magnetic signals, SenseMag classify the types of vehicles accordingly. Some semi-automated learning techniques have been adopted for the design of filters, features, and the choice of hyper-parameters. Extensive experiment based on real-word field deployment (on the highways in Shenzhen, China) shows that SenseMag significantly outperforms the existing methods in both classification accuracy and the granularity of vehicle types (i.e., 7 types by SenseMag versus 4 types by the existing work in comparisons). To be specific, our field experiment results validate that SenseMag is with at least $90\%$ vehicle type classification accuracy and less than 5\% vehicle length classification error.

AgFlow: Fast Model Selection of Penalized PCA via Implicit Regularization Effects of Gradient Flow

Principal component analysis (PCA) has been widely used as an effective technique for feature extraction and dimension reduction. In the High Dimension Low Sample Size (HDLSS) setting, one may prefer modified principal components, with penalized loadings, and automated penalty selection by implementing model selection among these different models with varying penalties. The earlier work [1, 2] has proposed penalized PCA, indicating the feasibility of model selection in $L_2$- penalized PCA through the solution path of Ridge regression, however, it is extremely time-consuming because of the intensive calculation of matrix inverse. In this paper, we propose a fast model selection method for penalized PCA, named Approximated Gradient Flow (AgFlow), which lowers the computation complexity through incorporating the implicit regularization effect introduced by (stochastic) gradient flow [3, 4] and obtains the complete solution path of $L_2$-penalized PCA under varying $L_2$-regularization. We perform extensive experiments on real-world datasets. AgFlow outperforms existing methods (Oja [5], Power [6], and Shamir [7] and the vanilla Ridge estimators) in terms of computation costs.

Exploring the Common Principal Subspace of Deep Features in Neural Networks

We find that different Deep Neural Networks (DNNs) trained with the same dataset share a common principal subspace in latent spaces, no matter in which architectures (e.g., Convolutional Neural Networks (CNNs), Multi-Layer Preceptors (MLPs) and Autoencoders (AEs)) the DNNs were built or even whether labels have been used in training (e.g., supervised, unsupervised, and self-supervised learning). Specifically, we design a new metric $\mathcal{P}$-vector to represent the principal subspace of deep features learned in a DNN, and propose to measure angles between the principal subspaces using $\mathcal{P}$-vectors. Small angles (with cosine close to $1.0$) have been found in the comparisons between any two DNNs trained with different algorithms/architectures. Furthermore, during the training procedure from random scratch, the angle decrease from a larger one ($70^\circ-80^\circ$ usually) to the small one, which coincides the progress of feature space learning from scratch to convergence. Then, we carry out case studies to measure the angle between the $\mathcal{P}$-vector and the principal subspace of training dataset, and connect such angle with generalization performance. Extensive experiments with practically-used Multi-Layer Perceptron (MLPs), AEs and CNNs for classification, image reconstruction, and self-supervised learning tasks on MNIST, CIFAR-10 and CIFAR-100 datasets have been done to support our claims with solid evidences. Interpretability of Deep Learning, Feature Learning, and Subspaces of Deep Features

Do What Nature Did To Us: Evolving Plastic Recurrent Neural Networks For Task Generalization

While artificial neural networks (ANNs) have been widely adopted in machine learning, researchers are increasingly obsessed by the gaps between ANNs and biological neural networks (BNNs). In this paper, we propose a framework named as Evolutionary Plastic Recurrent Neural Networks} (EPRNN). Inspired by BNN, EPRNN composes Evolution Strategies, Plasticity Rules, and Recursion-based Learning all in one meta learning framework for generalization to different tasks. More specifically, EPRNN incorporates with nested loops for meta learning -- an outer loop searches for optimal initial parameters of the neural network and learning rules; an inner loop adapts to specific tasks. In the inner loop of EPRNN, we effectively attain both long term memory and short term memory by forging plasticity with recursion-based learning mechanisms, both of which are believed to be responsible for memristance in BNNs. The inner-loop setting closely simulate that of BNNs, which neither query from any gradient oracle for optimization nor require the exact forms of learning objectives. To evaluate the performance of EPRNN, we carry out extensive experiments in two groups of tasks: Sequence Predicting, and Wheeled Robot Navigating. The experiment results demonstrate the unique advantage of EPRNN compared to state-of-the-arts based on plasticity and recursion while yielding comparably good performance against deep learning based approaches in the tasks. The experiment results suggest the potential of EPRNN to generalize to variety of tasks and encourage more efforts in plasticity and recursion based learning mechanisms.

AutoGCL: Automated Graph Contrastive Learning via Learnable View Generators

Contrastive learning has been widely applied to graph representation learning, where the view generators play a vital role in generating effective contrastive samples. Most of the existing contrastive learning methods employ pre-defined view generation methods, e.g., node drop or edge perturbation, which usually cannot adapt to input data or preserve the original semantic structures well. To address this issue, we propose a novel framework named Automated Graph Contrastive Learning (AutoGCL) in this paper. Specifically, AutoGCL employs a set of learnable graph view generators orchestrated by an auto augmentation strategy, where every graph view generator learns a probability distribution of graphs conditioned by the input. While the graph view generators in AutoGCL preserve the most representative structures of the original graph in generation of every contrastive sample, the auto augmentation learns policies to introduce adequate augmentation variances in the whole contrastive learning procedure. Furthermore, AutoGCL adopts a joint training strategy to train the learnable view generators, the graph encoder, and the classifier in an end-to-end manner, resulting in topological heterogeneity yet semantic similarity in the generation of contrastive samples. Extensive experiments on semi-supervised learning, unsupervised learning, and transfer learning demonstrate the superiority of our AutoGCL framework over the state-of-the-arts in graph contrastive learning. In addition, the visualization results further confirm that the learnable view generators can deliver more compact and semantically meaningful contrastive samples compared against the existing view generation methods.

Cross-Model Consensus of Explanations and Beyond for Image Classification Models: An Empirical Study

Existing interpretation algorithms have found that, even deep models make the same and right predictions on the same image, they might rely on different sets of input features for classification. However, among these sets of features, some common features might be used by the majority of models. In this paper, we are wondering what are the common features used by various models for classification and whether the models with better performance may favor those common features. For this purpose, our works uses an interpretation algorithm to attribute the importance of features (e.g., pixels or superpixels) as explanations, and proposes the cross-model consensus of explanations to capture the common features. Specifically, we first prepare a set of deep models as a committee, then deduce the explanation for every model, and obtain the consensus of explanations across the entire committee through voting. With the cross-model consensus of explanations, we conduct extensive experiments using 80+ models on 5 datasets/tasks. We find three interesting phenomena as follows: (1) the consensus obtained from image classification models is aligned with the ground truth of semantic segmentation; (2) we measure the similarity of the explanation result of each model in the committee to the consensus (namely consensus score), and find positive correlations between the consensus score and model performance; and (3) the consensus score coincidentally correlates to the interpretability.

Semi-Supervised Active Learning with Temporal Output Discrepancy

While deep learning succeeds in a wide range of tasks, it highly depends on the massive collection of annotated data which is expensive and time-consuming. To lower the cost of data annotation, active learning has been proposed to interactively query an oracle to annotate a small proportion of informative samples in an unlabeled dataset. Inspired by the fact that the samples with higher loss are usually more informative to the model than the samples with lower loss, in this paper we present a novel deep active learning approach that queries the oracle for data annotation when the unlabeled sample is believed to incorporate high loss. The core of our approach is a measurement Temporal Output Discrepancy (TOD) that estimates the sample loss by evaluating the discrepancy of outputs given by models at different optimization steps. Our theoretical investigation shows that TOD lower-bounds the accumulated sample loss thus it can be used to select informative unlabeled samples. On basis of TOD, we further develop an effective unlabeled data sampling strategy as well as an unsupervised learning criterion that enhances model performance by incorporating the unlabeled data. Due to the simplicity of TOD, our active learning approach is efficient, flexible, and task-agnostic. Extensive experimental results demonstrate that our approach achieves superior performances than the state-of-the-art active learning methods on image classification and semantic segmentation tasks.