Recent years have seen an increase in the development of large deep learning (DL) models, which makes training efficiency crucial. Common practice is struggling with the trade-off between usability and performance. On one hand, DL frameworks such as PyTorch use dynamic graphs to facilitate model developers at a price of sub-optimal model training performance. On the other hand, practitioners propose various approaches to improving the training efficiency by sacrificing some of the flexibility, ranging from making the graph static for more thorough optimization (e.g., XLA) to customizing optimization towards large-scale distributed training (e.g., DeepSpeed and Megatron-LM).

Multi-task model training has been adopted to enable a single deep neural network model (often a large language model) to handle multiple tasks (e.g., question answering and text summarization). Multi-task training commonly receives input sequences of highly different lengths due to the diverse contexts of different tasks. Padding (to the same sequence length) or packing (short examples into long sequences of the same length) is usually adopted to prepare input samples for model training, which is nonetheless not space or computation efficient. This paper proposes a dynamic micro-batching approach to tackle sequence length variation and enable efficient multi-task model training. We advocate pipelineparallel training of the large model with variable-length micro-batches, each of which potentially comprises a different number of samples. We optimize micro-batch construction using a dynamic programming-based approach, and handle micro-batch execution time variation through dynamic pipeline and communication scheduling, enabling highly efficient pipeline training. Extensive evaluation on the FLANv2 dataset demonstrates up to 4.39x higher training throughput when training T5, and 3.25x when training GPT, as compared with packing-based baselines. DynaPipe's source code is publicly available at https://github.com/awslabs/optimizing-multitask-training-through-dynamic-pipelines.

This work proposes POMP, a prompt pre-training method for vision-language models. Being memory and computation efficient, POMP enables the learned prompt to condense semantic information for a rich set of visual concepts with over twenty-thousand classes. Once pre-trained, the prompt with a strong transferable ability can be directly plugged into a variety of visual recognition tasks including image classification, semantic segmentation, and object detection, to boost recognition performances in a zero-shot manner. Empirical evaluation shows that POMP achieves state-of-the-art performances on 21 datasets, e.g., 67.0% average accuracy on 10 classification datasets (+3.1% compared to CoOp) and 84.4 hIoU on open-vocabulary Pascal VOC segmentation (+6.9 compared to ZSSeg).

Large deep learning models have recently garnered substantial attention from both academia and industry. Nonetheless, frequent failures are observed during large model training due to large-scale resources involved and extended training time. Existing solutions have significant failure recovery costs due to the severe restriction imposed by the bandwidth of remote storage in which they store checkpoints.

Adversarial training (AT) defends deep neural networks against adversarial attacks. One challenge that limits its practical application is the performance degradation on clean samples. A major bottleneck identified by previous works is the widely used batch normalization (BN), which struggles to model the different statistics of clean and adversarial training samples in AT. Although the dominant approach is to extend BN to capture this mixture of distribution, we propose to completely eliminate this bottleneck by removing all BN layers in AT. Our normalizer-free robust training (NoFrost) method extends recent advances in normalizer-free networks to AT for its unexplored advantage on handling the mixture distribution challenge. We show that NoFrost achieves adversarial robustness with only a minor sacrifice on clean sample accuracy. On ImageNet with ResNet50, NoFrost achieves $74.06%$ clean accuracy, which drops merely $2.00%$ from standard training. In contrast, BN-based AT obtains $59.28%$ clean accuracy, suffering a significant $16.78%$ drop from standard training. In addition, NoFrost achieves a $23.56%$ adversarial robustness against PGD attack, which improves the $13.57%$ robustness in BN-based AT. We observe better model smoothness and larger decision margins from NoFrost, which make the models less sensitive to input perturbations and thus more robust. Moreover, when incorporating more data augmentations into NoFrost, it achieves comprehensive robustness against multiple distribution shifts. Code and pre-trained models are public at https://github.com/amazon-research/normalizer-free-robust-training.

Existing out-of-distribution (OOD) detection methods are typically benchmarked on training sets with balanced class distributions. However, in real-world applications, it is common for the training sets to have long-tailed distributions. In this work, we first demonstrate that existing OOD detection methods commonly suffer from significant performance degradation when the training set is long-tail distributed. Through analysis, we posit that this is because the models struggle to distinguish the minority tail-class in-distribution samples, from the true OOD samples, making the tail classes more prone to be falsely detected as OOD. To solve this problem, we propose Partial and Asymmetric Supervised Contrastive Learning (PASCL), which explicitly encourages the model to distinguish between tail-class in-distribution samples and OOD samples. To further boost in-distribution classification accuracy, we propose Auxiliary Branch Finetuning, which uses two separate branches of BN and classification layers for anomaly detection and in-distribution classification, respectively. The intuition is that in-distribution and OOD anomaly data have different underlying distributions. Our method outperforms previous state-of-the-art method by $1.29%$, $1.45%$, $0.69%$ anomaly detection false positive rate (FPR) and $3.24%$, $4.06%$, $7.89%$ in-distribution classification accuracy on CIFAR10-LT, CIFAR100-LT, and ImageNet-LT, respectively. Code and pre-trained models are available at https://github.com/amazon-research/long-tailed-ood-detection.

We present results from a large-scale experiment on pretraining encoders with non-embedding parameter counts ranging from 700M to 9.3B, their subsequent distillation into smaller models ranging from 17M-170M parameters, and their application to the Natural Language Understanding (NLU) component of a virtual assistant system. Though we train using 70% spoken-form data, our teacher models perform comparably to XLM-R and mT5 when evaluated on the written-form Cross-lingual Natural Language Inference (XNLI) corpus. We perform a second stage of pretraining on our teacher models using in-domain data from our system, improving error rates by 3.86% relative for intent classification and 7.01% relative for slot filling. We find that even a 170M-parameter model distilled from our Stage 2 teacher model has 2.88% better intent classification and 7.69% better slot filling error rates when compared to the 2.3B-parameter teacher trained only on public data (Stage 1), emphasizing the importance of in-domain data for pretraining. When evaluated offline using labeled NLU data, our 17M-parameter Stage 2 distilled model outperforms both XLM-R Base (85M params) and DistillBERT (42M params) by 4.23% to 6.14%, respectively. Finally, we present results from a full virtual assistant experimentation platform, where we find that models trained using our pretraining and distillation pipeline outperform models distilled from 85M-parameter teachers by 3.74%-4.91% on an automatic measurement of full-system user dissatisfaction.

Existing general purpose frameworks for gigantic model training, i.e., dense models with billions of parameters, cannot scale efficiently on cloud environment with various networking conditions due to large communication overheads. In this paper, we propose MiCS, which Minimizes the Communication Scale to bring down communication overhead. Specifically, by decreasing the number of participants in a communication collective, MiCS can utilize heterogeneous network bandwidth, reduce network traffic over slower links, reduce the latency of communications for maintaining high network bandwidth utilization, and amortize expensive global gradient synchronization overhead. Our evaluation on AWS shows that the system throughput of MiCS is up to 2.89× that of the state-of-the-art large model training systems. MiCS achieves near-linear scaling efficiency, which is up to 1.27× that of DeepSpeed. MiCS allows us to train a proprietary model with 100 billion parameters on 512 GPUs with 99.4% weak-scaling efficiency, and it is able to saturate over 54.5% theoretical computation power of each GPU on a public cloud with less GPU memory and more restricted networks than DGX-A100 clusters.

The scalability of Distributed Stochastic Gradient Descent (SGD) is today limited by communication bottlenecks. We propose a novel SGD variant: \underline{C}ommunication-efficient \underline{S}GD with \underline{E}rror \underline{R}eset, or \underline{CSER}. The key idea in CSER is first a new technique called ``error reset'' that adapts arbitrary compressors for SGD, producing bifurcated local models with periodic reset of resulting local residual errors. Second we introduce partial synchronization for both the gradients and the models, leveraging advantages from them. We prove the convergence of CSER for smooth non-convex problems. Empirical results show that when combined with highly aggressive compressors, the CSER algorithms accelerate the distributed training by nearly $10\times$ for CIFAR-100, and by $4.5\times$ for ImageNet.

We present GluonCV and GluonNLP, the deep learning toolkits for computer vision and natural language processing based on Apache MXNet (incubating). These toolkits provide state-of-the-art pre-trained models, training scripts, and training logs, to facilitate rapid prototyping and promote reproducible research. We also provide modular APIs with flexible building blocks to enable efficient customization. Leveraging the MXNet ecosystem, the deep learning models in GluonCV and GluonNLP can be deployed onto a variety of platforms with different programming languages. The Apache 2.0 license has been adopted by GluonCV and GluonNLP to allow for software distribution, modification, and usage.

Communication overhead is a major bottleneck hampering the scalability of distributed machine learning systems. Recently, there has been a surge of interest in using gradient compression to improve the communication efficiency of distributed neural network training. Using 1-bit quantization, signSGD with majority vote achieves a 32x reduction in communication cost. However, its convergence is based on unrealistic assumptions and can diverge in practice. In this paper, we propose a general distributed compressed SGD with Nesterov's momentum. We consider two-way compression, which compresses the gradients both to and from workers. Convergence analysis on nonconvex problems for general gradient compressors is provided. By partitioning the gradient into blocks, a blockwise compressor is introduced such that each gradient block is compressed and transmitted in 1-bit format with a scaling factor, leading to a nearly 32x reduction on communication. Experimental results show that the proposed method converges as fast as full-precision distributed momentum SGD and achieves the same testing accuracy. In particular, on distributed ResNet training with 7 workers on the ImageNet, the proposed algorithm achieves the same testing accuracy as momentum SGD using full-precision gradients, but with $46\%$ less wall clock time.

In this paper, we introduce an attribute-based interactive image search which can leverage human-in-the-loop feedback to iteratively refine image search results. We study active image search where human feedback is solicited exclusively in visual form, without using relative attribute annotations used by prior work which are not typically found in many datasets. In order to optimize the image selection strategy, a deep reinforcement model is trained to learn what images are informative rather than rely on hand-crafted measures typically leveraged in prior work. Additionally, we extend the recently introduced Conditional Similarity Network to incorporate global similarity in training visual embeddings, which results in more natural transitions as the user explores the learned similarity embeddings. Our experiments demonstrate the effectiveness of our approach, producing compelling results on both active image search and image attribute representation tasks.

Variance reduction has been commonly used in stochastic optimization. It relies crucially on the assumption that the data set is finite. However, when the data are imputed with random noise as in data augmentation, the perturbed data set becomes essentially infinite. Recently, the stochastic MISO (S-MISO) algorithm is introduced to address this expected risk minimization problem. Though it converges faster than SGD, a significant amount of memory is required. In this paper, we propose two SGD-like algorithms for expected risk minimization with random perturbation, namely, stochastic sample average gradient (SSAG) and stochastic SAGA (S-SAGA). The memory cost of SSAG does not depend on the sample size, while that of S-SAGA is the same as those of variance reduction methods on unperturbed data. Theoretical analysis and experimental results on logistic regression and AUC maximization show that SSAG has faster convergence rate than SGD with comparable space requirement while S-SAGA outperforms S-MISO in terms of both iteration complexity and storage.

Understanding clothes from a single image would have huge commercial and cultural impacts on modern societies. However, this task remains a challenging computer vision problem due to wide variations in the appearance, style, brand and layering of clothing items. We present a new database called ModaNet, a large-scale collection of images based on Paperdoll dataset. Our dataset provides 55,176 street images, fully annotated with polygons on top of the 1 million weakly annotated street images in Paperdoll. ModaNet aims to provide a technical benchmark to fairly evaluate the progress of applying the latest computer vision techniques that rely on large data for fashion understanding. The rich annotation of the dataset allows to measure the performance of state-of-the-art algorithms for object detection, semantic segmentation and polygon prediction on street fashion images in detail.

Deep networks are highly nonlinear and difficult to optimize. During training, the parameter iterate may move from one local basin to another, or the data distribution may even change. Inspired by the close connection between stochastic optimization and online learning, we propose a variant of the follow the regularized leader (FTRL) algorithm called follow the moving leader (FTML). Unlike the FTRL family of algorithms, the recent samples are weighted more heavily in each iteration and so FTML can adapt more quickly to changes. We show that FTML enjoys the nice properties of RMSprop and Adam, while avoiding their pitfalls. Experimental results on a number of deep learning models and tasks demonstrate that FTML converges quickly, and outperforms other state-of-the-art optimizers.

With the recent proliferation of large-scale learning problems, there have been a lot of interest on distributed machine learning algorithms, particularly those that are based on stochastic gradient descent (SGD) and its variants. However, existing algorithms either suffer from slow convergence due to the inherent variance of stochastic gradients, or have a fast linear convergence rate but at the expense of poorer solution quality. In this paper, we combine their merits by proposing a fast distributed asynchronous SGD-based algorithm with variance reduction. A constant learning rate can be used, and it is also guaranteed to converge linearly to the optimal solution. Experiments on the Google Cloud Computing Platform demonstrate that the proposed algorithm outperforms state-of-the-art distributed asynchronous algorithms in terms of both wall clock time and solution quality.

In this demo paper we present a multiscale browsing interface for handheld devices, in which the user can interactively change the scale of the storyboard to easily adjust the amount of information desired. Conventional and hierarchical storyboards provide one or very few possible lengths. In contrast, scalable storyboards allow the number of images and the storyboard itself to be adapted to the device constraints (e.g. aspect ratio, resolution) and navigation state with much finer granularity. Several levels and modes, including segment of interest, are provided for more intuitive and convenient navigation.