Distributed training methods are crucial for large language models (LLMs). However, existing distributed training methods often suffer from communication bottlenecks, stragglers, and limited elasticity, particularly in heterogeneous or large-scale environments. Local SGD methods have been proposed to address these issues, but their effectiveness remains limited to small-scale training due to additional memory overhead and lack of concerns on efficiency and stability. To tackle these issues, we propose EDiT, an innovative Efficient Distributed Training method that combines a tailored Local SGD approach with model sharding techniques to enhance large-scale training efficiency. EDiT performs layer-wise parameter synchronization during forward pass, reducing communication and memory overhead and enabling overlap. Besides, EDiT employs a pseudo gradient penalty strategy to suppress loss spikes, which ensures training stability and improves performance. Additionally, we introduce A-EDiT, a fully asynchronous variant of EDiT that accommodates heterogeneous clusters. Building on EDiT/A-EDiT, we conduct a series of experiments to validate large-scale asynchronous training for LLMs, accompanied by comprehensive analyses. Experimental results demonstrate the superior performance of EDiT/A-EDiT, establishing them as robust solutions for distributed LLM training in diverse computational ecosystems. The code is available at Atorch codebase: https://github.com/intelligent-machine-learning/atorch/tree/main/atorch/local_sgd.

Vertical Federated Learning (VFL) is a solution increasingly used by companies with the same user group but differing features, enabling them to collaboratively train a machine learning model. VFL ensures that clients exchange intermediate results extracted by their local models, without sharing raw data. However, in practice, VFL encounters several challenges, such as computational and communication overhead, privacy leakage risk, and adversarial attack. Our study reveals that the usage of floating-point (FP) numbers is a common factor causing these issues, as they can be redundant and contain too much information. To address this, we propose a new architecture called rounding layer, which converts intermediate results to integers. Our theoretical analysis and empirical results demonstrate the benefits of the rounding layer in reducing computation and memory overhead, providing privacy protection, preserving model performance, and mitigating adversarial attacks. We hope this paper inspires further research into novel architectures to address practical issues in VFL.

Wildlife trafficking (WLT) has evolved into a pressing global concern, as traffickers increasingly utilize online platforms such as e-commerce websites and social networks to expand their illicit trade. This paper addresses the pivotal challenge of detecting and recognizing promotional behaviors related to the sale of wildlife products within online social networks—a critical step in combating these environmentally detrimental activities. To confront these illicit operations effectively, our research undertakes the following key initiatives: 1. Data Collection and Labeling: We employ a network-based approach to gather a scalable dataset pertaining to wildlife product trading. Through a human-in-the-loop machine learning process, this dataset is meticulously labeled, distinguishing between positive class samples containing wildlife product selling posts and hard-negatives representing regular posts misclassified as potential WLT posts, subsequently rectified by human annotators. 2. Machine Learning Framework Development: We present a robust framework that benchmarks machine learning results on the collected dataset. This framework autonomously identifies suspicious wildlife selling posts and accounts, effectively harnessing the multi-modal nature of online social networks. 3. In-depth Analysis of Trading Behaviors: Our research delves into a comprehensive analysis of trading posts, illuminating the systematic and organized selling behaviors prevalent in the current landscape. By providing detailed insights into the nature of these behaviors, we contribute valuable information for understanding and countering illegal wildlife product trading. Moreover, we emphasize our commitment to openness and collaboration by making our code and dataset openly available, thereby fostering cooperative efforts towards the development of more effective strategies in combating illegal wildlife trafficking.

In recent years, there has been a growing trend of incorporating hyperbolic geometry methods into computer vision. While these methods have achieved state-of-the-art performance on various metric learning tasks using hyperbolic distance measurements, the underlying theoretical analysis supporting this superior performance remains underexploited. In this study, we investigate the effects of integrating hyperbolic space into metric learning, particularly when training with contrastive loss. We identify a need for a comprehensive comparison between Euclidean and hyperbolic spaces regarding the temperature effect in the contrastive loss within the existing literature. To address this gap, we conduct an extensive investigation to benchmark the results of Vision Transformers (ViTs) using a hybrid objective function that combines loss from Euclidean and hyperbolic spaces. Additionally, we provide a theoretical analysis of the observed performance improvement. We also reveal that hyperbolic metric learning is highly related to hard negative sampling, providing insights for future work. This work will provide valuable data points and experience in understanding hyperbolic image embeddings. To shed more light on problem-solving and encourage further investigation into our approach, our code1is available online.

Chamfer distance (CD) is a standard metric to measure the shape dissimilarity between point clouds in point cloud completion, as well as a loss function for (deep) learning. However, it is well known that CD is vulnerable to outliers, leading to the drift towards suboptimal models. In contrast to the literature where most works address such issues in Euclidean space, we propose an extremely simple yet powerful metric for point cloud completion, namely Hyperbolic Chamfer Distance (HyperCD), that computes CD in hyperbolic space. In backpropagation, HyperCD consistently assigns higher weights to the matched point pairs with smaller Euclidean distances. In this way, good point matches are likely to be preserved while bad matches can be updated gradually, leading to better completion results. We demonstrate state-of-the-art performance on the benchmark datasets, i.e. PCN, ShapeNet-55, and ShapeNet34, and show from visualization that HyperCD can significantly improve the surface smoothness. Code is available at: https://github.com/Zhang-VISLab.

Deep Neural Networks (DNNs) generalization is known to be closely related to the flatness of minima, leading to the development of Sharpness-Aware Minimization (SAM) for seeking flatter minima and better generalization. In this paper, we revisit the loss of SAM and propose a more general method, called WSAM, by incorporating sharpness as a regularization term. We prove its generalization bound through the combination of PAC and Bayes-PAC techniques, and evaluate its performance on various public datasets. The results demonstrate that WSAM achieves improved generalization, or is at least highly competitive, compared to the vanilla optimizer, SAM and its variants. The code is available at this link https://github.com/intelligent-machine-learning/dlrover/tree/master/atorch/atorch/optimizers.

A point cloud is a discrete set of data points sampled from a 3D geometric surface. Chamfer distance (CD) is a popular metric and training loss to measure the distances between point clouds, but also well known to be sensitive to outliers. To address this issue, in this paper we propose InfoCD, a novel contrastive Chamfer distance loss to learn to spread the matched points for better distribution alignments between point clouds as well as accounting for a surface similarity estimator. We show that minimizing InfoCD is equivalent to maximizing a lower bound of the mutual information between the underlying geometric surfaces represented by the point clouds, leading to a regularized CD metric which is robust and computationally efficient for deep learning. We conduct comprehensive experiments for point cloud completion using InfoCD and observe significant improvements consistently over all the popular baseline networks trained with CD-based losses, leading to new state-of-the-art results on several benchmark datasets. Demo code is available at https://github.com/Zhang-VISLab/NeurIPS2023-InfoCD.

Adaptive optimizers, such as Adam, have achieved remarkable success in deep learning. A key component of these optimizers is the so-called preconditioning matrix, providing enhanced gradient information and regulating the step size of each gradient direction. In this paper, we propose a novel approach to designing the preconditioning matrix by utilizing the gradient difference between two successive steps as the diagonal elements. These diagonal elements are closely related to the Hessian and can be perceived as an approximation of the inner product between the Hessian row vectors and difference of the adjacent parameter vectors. Additionally, we introduce an auto-switching function that enables the preconditioning matrix to switch dynamically between Stochastic Gradient Descent (SGD) and the adaptive optimizer. Based on these two techniques, we develop a new optimizer named AGD that enhances the generalization performance. We evaluate AGD on public datasets of Natural Language Processing (NLP), Computer Vision (CV), and Recommendation Systems (RecSys). Our experimental results demonstrate that AGD outperforms the state-of-the-art (SOTA) optimizers, achieving highly competitive or significantly better predictive performance. Furthermore, we analyze how AGD is able to switch automatically between SGD and the adaptive optimizer and its actual effects on various scenarios. The code is available at https://github.com/intelligent-machine-learning/dlrover/tree/master/atorch/atorch/optimizers.

Data augmentation has been widely studied as it can be used to improve the generalizability of graph representation learning models. However, existing works focus only on the data augmentation on homogeneous graphs. Data augmentation for heterogeneous graphs remains under-explored. Considering that heterogeneous graphs contain different types of nodes and links, ignoring the type information and directly applying the data augmentation methods of homogeneous graphs to heterogeneous graphs will lead to suboptimal results. In this paper, we propose a novel Multi-Aspect Heterogeneous Graph Augmentation framework named MAHGA. Specifically, MAHGA consists of two core augmentation strategies: structure-level augmentation and metapath-level augmentation. Structure-level augmentation pays attention to network schema aspect and designs a relation-aware conditional variational auto-encoder that can generate synthetic features of neighbors to augment the nodes and the node types with scarce links. Metapath-level augmentation concentrates on metapath aspect, which constructs metapath reachable graphs for different metapaths and estimates the graphons of them. By sampling and mixing up based on the graphons, MAHGA yields intra-metapath and inter-metapath augmentation. Finally, we conduct extensive experiments on multiple benchmarks to validate the effectiveness of MAHGA. Experimental results demonstrate that our method improves the performances across a set of heterogeneous graph learning models and datasets.

With the recent advances in mobile sensing technologies, large amounts of sequential data are collected, such as vehicle GPS records, stock prices, sensor data from air quality detectors. Recurrent neural networks (RNNs) have been studied extensively to learn complex patterns for sequential data, with applicatons in natural language processing for sentence prediction/completion, human activity recognition for predicting or classifying human activities. However, there are many practical issues when training RNNs, e.g., vanishing and exploding gradients often occur due to the repeatability of network weights, etc. In this paper, we study the training stability in deep recurrent neural networks (RNNs), and propose a novel network, namely, deep incremental RNN (DIRNN). In contrast to the literature, we prove that DIRNN is essentially a Lyapunov stable dynamical system where there is no vanishing or exploding gradient in training. To demonstrate the applicability in practice, we also propose a novel implementation, namely TinyRNN, that sparsifies the transition matrices in DIRNN using weighted random permutations to reduce the model sizes. We evaluate our approach on seven benchmark datasets, and achieve state-of-the-art results. Demo code is provided in the supplementary file.

In this paper we consider the training stability of recurrent neural networks (RNNs) and propose a family of RNNs, namely SBO-RNN, that can be formulated using stochastic bilevel optimization (SBO). With the help of stochastic gradient descent (SGD), we manage to convert the SBO problem into an RNN where the feedforward and backpropagation solve the lower and upper-level optimization for learning hidden states and their hyperparameters, respectively. We prove that under mild conditions there is no vanishing or exploding gradient in training SBO-RNN. Empirically we demonstrate our approach with superior performance on several benchmark datasets, with fewer parameters, less training data, and much faster convergence. Code is available at https://zhang-vislab.github.io.

We propose a novel and efficient training method for RNNs by iteratively seeking a local minima on the loss surface within a small region, and leverage this directional vector for the update, in an outer-loop. We propose to utilize the Frank-Wolfe (FW) algorithm in this context. Although, FW implicitly involves normalized gradients, which can lead to a slow convergence rate, we develop a novel RNN training method that, surprisingly, even with the additional cost, the overall training cost is empirically observed to be lower than backpropagation. Our method leads to a new Frank-Wolfe method, that is in essence an SGD algorithm with a restart scheme. We prove that under certain conditions our algorithm has a sublinear convergence rate ofO(1/ϵ) for ϵ error. We then conduct empirical experiments on several benchmark datasets including those that exhibit long-term dependencies, and show significant performance improvement. We also experiment with deep RNN architectures and show efficient training performance. Finally, we demonstrate that our training method is robust to noisy data.