We now address major reviewer concerns below. Our work's goal is to identify statistics that will allow identification of We've obtained results for shallow architectures with CIFAR-10 dataset - biologically, perhaps Will include these results in revision. We will improve discussion of this in revision. Schrimpf et al. 2019] - just not at such large scale as the included results, since such networks are very resource-36 However, outcomes don't change conclusions at all, will include what we have in revision. We found largely consistent results (Fig. S4).

Large-scale deep neural networks (DNN) exhibit excellent performance for various tasks. As DNNs and datasets grow, distributed training becomes extremely time-consuming and demands larger clusters. A main bottleneck is the resulting gradient aggregation overhead. While gradient compression and sparse collective communication techniques are commonly employed to alleviate network load, many gradient compression schemes do not achieve acceleration of the training process while also preserving accuracy. This paper introduces PacTrain, a novel framework that accelerates distributed training by combining pruning with sparse gradient compression. Active pruning of the neural network makes the model weights and gradients sparse. By ensuring the global knowledge of the gradient sparsity among all distributed training workers, we can perform lightweight compression communication without harming accuracy. We show that the PacTrain compression scheme achieves a near-optimal compression strategy while remaining compatible with the all-reduce primitive. Experimental evaluations show that PacTrain improves training throughput by 1.25 to 8.72 times compared to state-of-the-art compression-enabled systems for representative vision and language models training tasks under bandwidth-constrained conditions.

Distributed training is the de facto standard to scale up the training of Deep Neural Networks (DNNs) with multiple GPUs. The performance bottleneck of distributed training lies in communications for gradient synchronization. Recently, practitioners have observed sparsity in gradient tensors, suggesting the potential to reduce the traffic volume in communication and improve end-to-end training efficiency. Yet, the optimal communication scheme to fully leverage sparsity is still missing. This paper aims to address this gap. We first analyze the characteristics of sparse tensors in popular DNN models to understand the fundamentals of sparsity. We then systematically explore the design space of communication schemes for sparse tensors and find the optimal one. % We then find the optimal scheme based on the characteristics by systematically exploring the design space. We also develop a gradient synchronization system called Zen that approximately realizes it for sparse tensors. We demonstrate that Zen can achieve up to 5.09x speedup in communication time and up to 2.48x speedup in training throughput compared to the state-of-the-art methods.

Large Language Models (LLMs) have revolutionized Natural Language Processing (NLP) but demand massive GPU resources for training. Lowering the threshold for LLMs training would encourage greater participation from researchers, benefiting both academia and society. While existing approaches have focused on parameter-efficient fine-tuning, which tunes or adds a small number of parameters, few have addressed the challenge of tuning the full parameters of LLMs with limited resources. In this work, we propose a new optimizer, LOw-Memory Optimization (LOMO), which fuses the gradient computation and the parameter update in one step to reduce memory usage. By integrating LOMO with existing memory saving techniques, we reduce memory usage to 10.8% compared to the standard approach (DeepSpeed solution). Consequently, our approach enables the full parameter fine-tuning of a 65B model on a single machine with 8 RTX 3090, each with 24GB memory.

The tensor data recovery task has thus attracted much research attention in recent years. Solving such an ill-posed problem generally requires to explore intrinsic prior structures underlying tensor data, and formulate them as certain forms of regularization terms for guiding a sound estimate of the restored tensor. Recent research have made significant progress by adopting two insightful tensor priors, i.e., global low-rankness (L) and local smoothness (S) across different tensor modes, which are always encoded as a sum of two separate regularization terms into the recovery models. However, unlike the primary theoretical developments on low-rank tensor recovery, these joint L+S models have no theoretical exact-recovery guarantees yet, making the methods lack reliability in real practice. To this crucial issue, in this work, we build a unique regularization term, which essentially encodes both L and S priors of a tensor simultaneously. Especially, by equipping this single regularizer into the recovery models, we can rigorously prove the exact recovery guarantees for two typical tensor recovery tasks, i.e., tensor completion (TC) and tensor robust principal component analysis (TRPCA). To the best of our knowledge, this should be the first exact-recovery results among all related L+S methods for tensor recovery. Significant recovery accuracy improvements over many other SOTA methods in several TC and TRPCA tasks with various kinds of visual tensor data are observed in extensive experiments. Typically, our method achieves a workable performance when the missing rate is extremely large, e.g., 99.5%, for the color image inpainting task, while all its peers totally fail in such challenging case.

Training deep neural networks on large datasets containing high-dimensional data requires a large amount of computation. A solution to this problem is data-parallel distributed training, where a model is replicated into several computational nodes that have access to different chunks of the data. This approach, however, entails high communication rates and latency because of the computed gradients that need to be shared among nodes at every iteration. The problem becomes more pronounced in the case that there is wireless communication between the nodes (i.e. due to the limited network bandwidth). To address this problem, various compression methods have been proposed including sparsification, quantization, and entropy encoding of the gradients. Existing methods leverage the intra-node information redundancy, that is, they compress gradients at each node independently. In contrast, we advocate that the gradients across the nodes are correlated and propose methods to leverage this inter-node redundancy to improve compression efficiency. Depending on the node communication protocol (parameter server or ring-allreduce), we propose two instances of the LGC approach that we coin Learned Gradient Compression (LGC). Our methods exploit an autoencoder (i.e. trained during the first stages of the distributed training) to capture the common information that exists in the gradients of the distributed nodes. We have tested our LGC methods on the image classification and semantic segmentation tasks using different convolutional neural networks (ResNet50, ResNet101, PSPNet) and multiple datasets (ImageNet, Cifar10, CamVid). The ResNet101 model trained for image classification on Cifar10 achieved an accuracy of 93.57%, which is lower than the baseline distributed training with uncompressed gradients only by 0.18%.

We propose gradient adversarial training, an auxiliary deep learning framework applicable to different machine learning problems. In gradient adversarial training, we leverage a prior belief that in many contexts, simultaneous gradient updates should be statistically indistinguishable from each other. We enforce this consistency using an auxiliary network that classifies the origin of the gradient tensor, and the main network serves as an adversary to the auxiliary network in addition to performing standard task-based training. We demonstrate gradient adversarial training for three different scenarios: (1) as a defense to adversarial examples we classify gradient tensors and tune them to be agnostic to the class of their corresponding example, (2) for knowledge distillation, we do binary classification of gradient tensors derived from the student or teacher network and tune the student gradient tensor to mimic the teacher's gradient tensor; and (3) for multi-task learning we classify the gradient tensors derived from different task loss functions and tune them to be statistically indistinguishable. For each of the three scenarios we show the potential of gradient adversarial training procedure. Specifically, gradient adversarial training increases the robustness of a network to adversarial attacks, is able to better distill the knowledge from a teacher network to a student network compared to soft targets, and boosts multi-task learning by aligning the gradient tensors derived from the task specific loss functions. Overall, our experiments demonstrate that gradient tensors contain latent information about whatever tasks are being trained, and can support diverse machine learning problems when intelligently guided through adversarialization using a auxiliary network.