Search

Distributed learning is indispensable for training large-scale deep models. Asynchronous SGD~(ASGD) and its variants are commonly used distributed learning methods in many scenarios where the computing capabilities of workers in the cluster are heterogeneous. Momentum has been acknowledged for its benefits in both optimization and generalization in deep model training. However, existing works have found that naively incorporating momentum into ASGD can impede the convergence. In this paper, we propose a novel method, called ordered momentum (OrMo), for ASGD. In OrMo, momentum is incorporated into ASGD by organizing the gradients in order based on their iteration indexes. We theoretically prove the convergence of OrMo for non-convex problems. To the best of our knowledge, this is the first work to establish the convergence analysis of ASGD with momentum without relying on the bounded delay assumption. Empirical results demonstrate that OrMo can achieve better convergence performance compared with ASGD and other asynchronous methods with momentum.

Asynchronous federated learning (AFL) is an effective method to address the challenge of device heterogeneity in cross-device federated learning. However, AFL is usually incompatible with existing secure aggregation protocols used to protect user privacy in federated learning because most existing secure aggregation protocols are based on synchronous aggregation. To address this problem, we propose a novel secure aggregation protocol named buffered asynchronous secure aggregation (BASA) in this paper. Compared with existing protocols, BASA is fully compatible with AFL and provides secure aggregation under the condition that each user only needs one round of communication with the server without relying on any synchronous interaction among users. Based on BASA, we propose the first AFL method which achieves secure aggregation without extra requirements on hardware. We empirically demonstrate that BASA outperforms existing secure aggregation protocols for cross-device federated learning in terms of training efficiency and scalability.

Large language models~(LLMs) have recently demonstrated promising performance in many tasks. However, the high storage and computational cost of LLMs has become a challenge for deploying LLMs. Weight quantization has been widely used for model compression, which can reduce both storage and computational cost. Most existing weight quantization methods for LLMs use a rank-one codebook for quantization, which results in substantial accuracy loss when the compression ratio is high. In this paper, we propose a novel weight quantization method, called low-rank codebook based quantization~(LCQ), for LLMs. LCQ adopts a low-rank codebook, the rank of which can be larger than one, for quantization. Experiments show that LCQ can achieve better accuracy than existing methods with a negligibly extra storage cost., Comment: 10 pages, 5 figures

Continual learning requires the model to learn multiple tasks sequentially. In continual learning, the model should possess the ability to maintain its performance on old tasks (stability) and the ability to adapt to new tasks continuously (plasticity). Recently, parameter-efficient fine-tuning (PEFT), which involves freezing a pre-trained model and injecting a small number of learnable parameters to adapt to downstream tasks, has gained increasing popularity in continual learning. Although existing continual learning methods based on PEFT have demonstrated superior performance compared to those not based on PEFT, most of them do not consider how to eliminate the interference of the new task on the old tasks, which inhibits the model from making a good trade-off between stability and plasticity. In this work, we propose a new PEFT method, called interference-free low-rank adaptation (InfLoRA), for continual learning. InfLoRA injects a small number of parameters to reparameterize the pre-trained weights and shows that fine-tuning these injected parameters is equivalent to fine-tuning the pre-trained weights within a subspace. Furthermore, InfLoRA designs this subspace to eliminate the interference of the new task on the old tasks, making a good trade-off between stability and plasticity. Experimental results show that InfLoRA outperforms existing state-of-the-art continual learning methods on multiple datasets., Comment: Accepted by the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2024)

Distributed learning is commonly used for training deep learning models, especially large models. In distributed learning, manual parallelism (MP) methods demand considerable human effort and have limited flexibility. Hence, automatic parallelism (AP) methods have recently been proposed for automating the parallel strategy optimization process. Existing AP methods suffer from sub-optimal solutions because they do not jointly optimize the two categories of parallel strategies (i.e., inter-layer parallelism and intra-layer parallelism). In this paper, we propose a novel AP method called UniAP, which unifies inter- and intra-layer automatic parallelism by mixed integer quadratic programming. To the best of our knowledge, UniAP is the first parallel method that can jointly optimize the two categories of parallel strategies to find an optimal solution. Experimental results show that UniAP outperforms state-of-the-art methods by up to 3.80$\times$ in throughput and reduces strategy optimization time by up to 107$\times$ across five Transformer-based models., Comment: 20 pages, 8 figures

Byzantine-robust distributed learning (BRDL), in which computing devices are likely to behave abnormally due to accidental failures or malicious attacks, has recently become a hot research topic. However, even in the independent and identically distributed (i.i.d.) case, existing BRDL methods will suffer from a significant drop on model accuracy due to the large variance of stochastic gradients. Increasing batch sizes is a simple yet effective way to reduce the variance. However, when the total number of gradient computation is fixed, a too-large batch size will lead to a too-small iteration number (update number), which may also degrade the model accuracy. In view of this challenge, we mainly study the optimal batch size when the total number of gradient computation is fixed in this work. In particular, we theoretically and empirically show that when the total number of gradient computation is fixed, the optimal batch size in BRDL increases with the fraction of Byzantine workers. Therefore, compared to the case without attacks, the batch size should be set larger when under Byzantine attacks. However, for existing BRDL methods, large batch sizes will lead to a drop on model accuracy, even if there is no Byzantine attack. To deal with this problem, we propose a novel BRDL method, called Byzantine-robust stochastic gradient descent with normalized momentum (ByzSGDnm), which can alleviate the drop on model accuracy in large-batch cases. Moreover, we theoretically prove the convergence of ByzSGDnm for general non-convex cases under Byzantine attacks. Empirical results show that ByzSGDnm has a comparable performance to existing BRDL methods under bit-flipping failure, but can outperform existing BRDL methods under deliberately crafted attacks.

Federated learning (FL) has recently become a hot research topic, in which Byzantine robustness, communication efficiency and privacy preservation are three important aspects. However, the tension among these three aspects makes it hard to simultaneously take all of them into account. In view of this challenge, we theoretically analyze the conditions that a communication compression method should satisfy to be compatible with existing Byzantine-robust methods and privacy-preserving methods. Motivated by the analysis results, we propose a novel communication compression method called consensus sparsification (ConSpar). To the best of our knowledge, ConSpar is the first communication compression method that is designed to be compatible with both Byzantine-robust methods and privacy-preserving methods. Based on ConSpar, we further propose a novel FL framework called FedREP, which is Byzantine-robust, communication-efficient and privacy-preserving. We theoretically prove the Byzantine robustness and the convergence of FedREP. Empirical results show that FedREP can significantly outperform communication-efficient privacy-preserving baselines. Furthermore, compared with Byzantine-robust communication-efficient baselines, FedREP can achieve comparable accuracy with the extra advantage of privacy preservation.

Link prediction is a fundamental problem in many graph based applications, such as protein-protein interaction prediction. Graph neural network (GNN) has recently been widely used for link prediction. However, existing GNN based link prediction (GNN-LP) methods suffer from scalability problem during training for large-scale graphs, which has received little attention by researchers. In this paper, we first give computation complexity analysis of existing GNN-LP methods, which reveals that the scalability problem stems from their symmetric learning strategy adopting the same class of GNN models to learn representation for both head and tail nodes. Then we propose a novel method, called asymmetric learning (AML), for GNN-LP. The main idea of AML is to adopt a GNN model for learning head node representation while using a multi-layer perceptron (MLP) model for learning tail node representation. Furthermore, AML proposes a row-wise sampling strategy to generate mini-batch for training, which is a necessary component to make the asymmetric learning strategy work for training speedup. To the best of our knowledge, AML is the first GNN-LP method adopting an asymmetric learning strategy for node representation learning. Experiments on three real large-scale datasets show that AML is 1.7X~7.3X faster in training than baselines with a symmetric learning strategy, while having almost no accuracy loss.

Online continual learning, especially when task identities and task boundaries are unavailable, is a challenging continual learning setting. One representative kind of methods for online continual learning is replay-based methods, in which a replay buffer called memory is maintained to keep a small part of past samples for overcoming catastrophic forgetting. When tackling with online continual learning, most existing replay-based methods focus on single-label problems in which each sample in the data stream has only one label. But multi-label problems may also happen in the online continual learning setting in which each sample may have more than one label. In the online setting with multi-label samples, the class distribution in data stream is typically highly imbalanced, and it is challenging to control class distribution in memory since changing the number of samples belonging to one class may affect the number of samples belonging to other classes. But class distribution in memory is critical for replay-based memory to get good performance, especially when the class distribution in data stream is highly imbalanced. In this paper, we propose a simple but effective method, called optimizing class distribution in memory (OCDM), for multi-label online continual learning. OCDM formulates the memory update mechanism as an optimization problem and updates the memory by solving this problem. Experiments on two widely used multi-label datasets show that OCDM can control the class distribution in memory well and can outperform other state-of-the-art methods.

Multi-agent reinforcement learning (MARL) algorithms have made promising progress in recent years by leveraging the centralized training and decentralized execution (CTDE) paradigm. However, existing MARL algorithms still suffer from the sample inefficiency problem. In this paper, we propose a simple yet effective approach, called state-based episodic memory (SEM), to improve sample efficiency in MARL. SEM adopts episodic memory (EM) to supervise the centralized training procedure of CTDE in MARL. To the best of our knowledge, SEM is the first work to introduce EM into MARL. We can theoretically prove that, when using for MARL, SEM has lower space complexity and time complexity than state and action based EM (SAEM), which is originally proposed for single-agent reinforcement learning. Experimental results on StarCraft multi-agent challenge (SMAC) show that introducing episodic memory into MARL can improve sample efficiency and SEM can reduce storage cost and time cost compared with SAEM.

Animals are able to imitate each others' behavior, despite their difference in biomechanics. In contrast, imitating the other similar robots is a much more challenging task in robotics. This problem is called cross domain imitation learning~(CDIL). In this paper, we consider CDIL on a class of similar robots. We tackle this problem by introducing an imitation learning algorithm based on invariant representation. We propose to learn invariant state and action representations, which aligns the behavior of multiple robots so that CDIL becomes possible. Compared with previous invariant representation learning methods for similar purpose, our method does not require human-labeled pairwise data for training. Instead, we use cycle-consistency and domain confusion to align the representation and increase its robustness. We test the algorithm on multiple robots in simulator and show that unseen new robot instances can be trained with existing expert demonstrations successfully. Qualitative results also demonstrate that the proposed method is able to learn similar representations for different robots with similar behaviors, which is essential for successful CDIL.

Due to its low storage cost and fast query speed, hashing has been widely used in large-scale image retrieval tasks. Hash bucket search returns data points within a given Hamming radius to each query, which can enable search at a constant or sub-linear time cost. However, existing hashing methods cannot achieve satisfactory retrieval performance for hash bucket search in complex scenarios, since they learn only one hash code for each image. More specifically, by using one hash code to represent one image, existing methods might fail to put similar image pairs to the buckets with a small Hamming distance to the query when the semantic information of images is complex. As a result, a large number of hash buckets need to be visited for retrieving similar images, based on the learned codes. This will deteriorate the efficiency of hash bucket search. In this paper, we propose a novel hashing framework, called multiple code hashing (MCH), to improve the performance of hash bucket search. The main idea of MCH is to learn multiple hash codes for each image, with each code representing a different region of the image. Furthermore, we propose a deep reinforcement learning algorithm to learn the parameters in MCH. To the best of our knowledge, this is the first work that proposes to learn multiple hash codes for each image in image retrieval. Experiments demonstrate that MCH can achieve a significant improvement in hash bucket search, compared with existing methods that learn only one hash code for each image., Comment: 12 pages, 9 figures, 3 tables

Stochastic gradient descent~(SGD) and its variants have been the dominating optimization methods in machine learning. Compared to SGD with small-batch training, SGD with large-batch training can better utilize the computational power of current multi-core systems such as graphics processing units~(GPUs) and can reduce the number of communication rounds in distributed training settings. Thus, SGD with large-batch training has attracted considerable attention. However, existing empirical results showed that large-batch training typically leads to a drop in generalization accuracy. Hence, how to guarantee the generalization ability in large-batch training becomes a challenging task. In this paper, we propose a simple yet effective method, called stochastic normalized gradient descent with momentum~(SNGM), for large-batch training. We prove that with the same number of gradient computations, SNGM can adopt a larger batch size than momentum SGD~(MSGD), which is one of the most widely used variants of SGD, to converge to an $\epsilon$-stationary point. Empirical results on deep learning verify that when adopting the same large batch size, SNGM can achieve better test accuracy than MSGD and other state-of-the-art large-batch training methods.

Animals are able to discover the topological map (graph) of surrounding environment, which will be used for navigation. Inspired by this biological phenomenon, researchers have recently proposed to generate graph representation for Markov decision process (MDP) and use such graphs for planning in reinforcement learning (RL). However, existing graph generation methods suffer from many drawbacks. One drawback is that existing methods do not learn an abstraction for graphs, which results in high memory and computation cost. This drawback also makes generated graph non-robust, which degrades the planning performance. Another drawback is that existing methods cannot be used for facilitating exploration which is important in RL. In this paper, we propose a new method, called topological map abstraction (TOMA), for graph generation. TOMA can generate an abstract graph representation for MDP, which costs much less memory and computation cost than existing methods. Furthermore, TOMA can be used for facilitating exploration. In particular, we propose planning to explore, in which TOMA is used to accelerate exploration by guiding the agent towards unexplored states. A novel experience replay module called vertex memory is also proposed to improve exploration performance. Experimental results show that TOMA can outperform existing methods to achieve the state-of-the-art performance.

Distributed learning has become a hot research topic due to its wide application in clusterbased large-scale learning, federated learning, edge computing and so on. Most traditional distributed learning methods typically assume no failure or attack. However, many unexpected cases, such as communication failure and even malicious attack, may happen in real applications. Hence, Byzantine learning (BL), which refers to distributed learning with failure or attack, has recently attracted much attention. Most existing BL methods are synchronous, which are impractical in some applications due to heterogeneous or offline workers. In these cases, asynchronous BL (ABL) is usually preferred. In this paper, we propose a novel method, called buffered asynchronous stochastic gradient descent (BASGD), for ABL. To the best of our knowledge, BASGD is the first ABL method that can resist non-omniscient attacks without storing any instances on server. Furthermore, we also propose an improved variant of BASGD, called BASGD with momentum (BASGDm), by introducing momentum into BASGD. BASGDm can resist both non-omniscient and omniscient attacks. Compared with those methods which need to store instances on server, BASGD and BASGDm have a wider scope of application. Both BASGD and BASGDm are compatible with various aggregation rules. Moreover, both BASGD and BASGDm are proved to be convergent and be able to resist failure or attack. Empirical results show that our methods significantly outperform existing ABL baselines when there exists failure or attack on workers.

Existing research shows that the batch size can seriously affect the performance of stochastic gradient descent~(SGD) based learning, including training speed and generalization ability. A larger batch size typically results in less parameter updates. In distributed training, a larger batch size also results in less frequent communication. However, a larger batch size can make a generalization gap more easily. Hence, how to set a proper batch size for SGD has recently attracted much attention. Although some methods about setting batch size have been proposed, the batch size problem has still not been well solved. In this paper, we first provide theory to show that a proper batch size is related to the gap between initialization and optimum of the model parameter. Then based on this theory, we propose a novel method, called \underline{s}tagewise \underline{e}nlargement of \underline{b}atch \underline{s}ize~(\mbox{SEBS}), to set proper batch size for SGD. More specifically, \mbox{SEBS} adopts a multi-stage scheme, and enlarges the batch size geometrically by stage. We theoretically prove that, compared to classical stagewise SGD which decreases learning rate by stage, \mbox{SEBS} can reduce the number of parameter updates without increasing generalization error. SEBS is suitable for \mbox{SGD}, momentum \mbox{SGD} and AdaGrad. Empirical results on real data successfully verify the theories of \mbox{SEBS}. Furthermore, empirical results also show that SEBS can outperform other baselines.

With the development of deep neural networks, the size of network models becomes larger and larger. Model compression has become an urgent need for deploying these network models to mobile or embedded devices. Model quantization is a representative model compression technique. Although a lot of quantization methods have been proposed, many of them suffer from a high quantization error caused by a long-tail distribution of network weights. In this paper, we propose a novel quantization method, called weight normalization based quantization (WNQ), for model compression. WNQ adopts weight normalization to avoid the long-tail distribution of network weights and subsequently reduces the quantization error. Experiments on CIFAR-100 and ImageNet show that WNQ can outperform other baselines to achieve state-of-the-art performance., Comment: 10 pages, 5 figures

Stochastic gradient decent~(SGD) and its variants, including some accelerated variants, have become popular for training in machine learning. However, in all existing SGD and its variants, the sample size in each iteration~(epoch) of training is the same as the size of the full training set. In this paper, we propose a new method, called \underline{ada}ptive \underline{s}ample \underline{s}election~(ADASS), for training acceleration. During different epoches of training, ADASS only need to visit different training subsets which are adaptively selected from the full training set according to the Lipschitz constants of the loss functions on samples. It means that in ADASS the sample size in each epoch of training can be smaller than the size of the full training set, by discarding some samples. ADASS can be seamlessly integrated with existing optimization methods, such as SGD and momentum SGD, for training acceleration. Theoretical results show that the learning accuracy of ADASS is comparable to that of counterparts with full training set. Furthermore, empirical results on both shallow models and deep models also show that ADASS can accelerate the training process of existing methods without sacrificing accuracy.

Exploration strategy design is one of the challenging problems in reinforcement learning~(RL), especially when the environment contains a large state space or sparse rewards. During exploration, the agent tries to discover novel areas or high reward~(quality) areas. In most existing methods, the novelty and quality in the neighboring area of the current state are not well utilized to guide the exploration of the agent. To tackle this problem, we propose a novel RL framework, called \underline{c}lustered \underline{r}einforcement \underline{l}earning~(CRL), for efficient exploration in RL. CRL adopts clustering to divide the collected states into several clusters, based on which a bonus reward reflecting both novelty and quality in the neighboring area~(cluster) of the current state is given to the agent. Experiments on a continuous control task and several \emph{Atari 2600} games show that CRL can outperform other state-of-the-art methods to achieve the best performance in most cases., Comment: 16pages, 3 figures

Distributed stochastic gradient descent~(DSGD) has been widely used for optimizing large-scale machine learning models, including both convex and non-convex models. With the rapid growth of model size, huge communication cost has been the bottleneck of traditional DSGD. Recently, many communication compression methods have been proposed. Memory-based distributed stochastic gradient descent~(M-DSGD) is one of the efficient methods since each worker communicates a sparse vector in each iteration so that the communication cost is small. Recent works propose the convergence rate of M-DSGD when it adopts vanilla SGD. However, there is still a lack of convergence theory for M-DSGD when it adopts momentum SGD. In this paper, we propose a universal convergence analysis for M-DSGD by introducing \emph{transformation equation}. The transformation equation describes the relation between traditional DSGD and M-DSGD so that we can transform M-DSGD to its corresponding DSGD. Hence we get the convergence rate of M-DSGD with momentum for both convex and non-convex problems. Furthermore, we combine M-DSGD and stagewise learning that the learning rate of M-DSGD in each stage is a constant and is decreased by stage, instead of iteration. Using the transformation equation, we propose the convergence rate of stagewise M-DSGD which bridges the gap between theory and practice.

With the rapid growth of data, distributed momentum stochastic gradient descent~(DMSGD) has been widely used in distributed learning, especially for training large-scale deep models. Due to the latency and limited bandwidth of the network, communication has become the bottleneck of distributed learning. Communication compression with sparsified gradient, abbreviated as \emph{sparse communication}, has been widely employed to reduce communication cost. All existing works about sparse communication in DMSGD employ local momentum, in which the momentum only accumulates stochastic gradients computed by each worker locally. In this paper, we propose a novel method, called \emph{\underline{g}}lobal \emph{\underline{m}}omentum \emph{\underline{c}}ompression~(GMC), for sparse communication. Different from existing works that utilize local momentum, GMC utilizes global momentum. Furthermore, to enhance the convergence performance when using more aggressive sparsification compressors (e.g., RBGS), we extend GMC to GMC+. We theoretically prove the convergence of GMC and GMC+. To the best of our knowledge, this is the first work that introduces global momentum for sparse communication in distributed learning. Empirical results demonstrate that, compared with the local momentum counterparts, our GMC and GMC+ can achieve higher test accuracy and exhibit faster convergence, especially under non-IID data distribution.

Recommender systems (RS), which have been an essential part in a wide range of applications, can be formulated as a matrix completion (MC) problem. To boost the performance of MC, matrix completion with side information, called inductive matrix completion (IMC), was further proposed. In real applications, the factorized version of IMC is more favored due to its efficiency of optimization and implementation. Regarding the factorized version, traditional IMC method can be interpreted as learning an individual representation for each feature, which is independent from each other. Moreover, representations for the same features are shared across all users/items. However, the independent characteristic for features and shared characteristic for the same features across all users/items may limit the expressiveness of the model. The limitation also exists in variants of IMC, such as deep learning based IMC models. To break the limitation, we generalize recent advances of self-attention mechanism to IMC and propose a context-aware model called collaborative self-attention (CSA), which can jointly learn context-aware representations for features and perform inductive matrix completion process. Extensive experiments on three large-scale datasets from real RS applications demonstrate effectiveness of CSA., Comment: There are large modifications

Due to its efficiency and ease to implement, stochastic gradient descent (SGD) has been widely used in machine learning. In particular, SGD is one of the most popular optimization methods for distributed learning. Recently, quantized SGD (QSGD), which adopts quantization to reduce the communication cost in SGD-based distributed learning, has attracted much attention. Although several QSGD methods have been proposed, some of them are heuristic without theoretical guarantee, and others have high quantization variance which makes the convergence become slow. In this paper, we propose a new method, called Quantized Epoch-SGD (QESGD), for communication-efficient distributed learning. QESGD compresses (quantizes) the parameter with variance reduction, so that it can get almost the same performance as that of SGD with less communication cost. QESGD is implemented on the Parameter Server framework, and empirical results on distributed deep learning show that QESGD can outperform other state-of-the-art quantization methods to achieve the best performance.

Distributed sparse learning with a cluster of multiple machines has attracted much attention in machine learning, especially for large-scale applications with high-dimensional data. One popular way to implement sparse learning is to use $L_1$ regularization. In this paper, we propose a novel method, called proximal \mbox{SCOPE}~(\mbox{pSCOPE}), for distributed sparse learning with $L_1$ regularization. pSCOPE is based on a \underline{c}ooperative \underline{a}utonomous \underline{l}ocal \underline{l}earning~(\mbox{CALL}) framework. In the \mbox{CALL} framework of \mbox{pSCOPE}, we find that the data partition affects the convergence of the learning procedure, and subsequently we define a metric to measure the goodness of a data partition. Based on the defined metric, we theoretically prove that pSCOPE is convergent with a linear convergence rate if the data partition is good enough. We also prove that better data partition implies faster convergence rate. Furthermore, pSCOPE is also communication efficient. Experimental results on real data sets show that pSCOPE can outperform other state-of-the-art distributed methods for sparse learning.

Geometric matrix completion (GMC) has been proposed for recommendation by integrating the relationship (link) graphs among users/items into matrix completion (MC). Traditional GMC methods typically adopt graph regularization to impose smoothness priors for MC. Recently, geometric deep learning on graphs (GDLG) is proposed to solve the GMC problem, showing better performance than existing GMC methods including traditional graph regularization based methods. To the best of our knowledge, there exists only one GDLG method for GMC, which is called RMGCNN. RMGCNN combines graph convolutional network (GCN) and recurrent neural network (RNN) together for GMC. In the original work of RMGCNN, RMGCNN demonstrates better performance than pure GCN-based method. In this paper, we propose a new GMC method, called convolutional geometric matrix completion (CGMC), for recommendation with graphs among users/items. CGMC is a pure GCN-based method with a newly designed graph convolutional network. Experimental results on real datasets show that CGMC can outperform other state-of-the-art methods including RMGCNN in terms of both accuracy and speed.

The frequent directions (FD) technique is a deterministic approach for online sketching that has many applications in machine learning. The conventional FD is a heuristic procedure that often outputs rank deficient matrices. To overcome the rank deficiency problem, we propose a new sketching strategy called robust frequent directions (RFD) by introducing a regularization term. RFD can be derived from an optimization problem. It updates the sketch matrix and the regularization term adaptively and jointly. RFD reduces the approximation error of FD without increasing the computational cost. We also apply RFD to online learning and propose an effective hyperparameter-free online Newton algorithm. We derive a regret bound for our online Newton algorithm based on RFD, which guarantees the robustness of the algorithm. The experimental studies demonstrate that the proposed method outperforms state-of-the-art second order online learning algorithms., Comment: Journal of Machine Learning Research 2019, Accepted