Distributed training often suffers from high communication overhead due to large-scale gradient synchronization. Although gradient compression—particularly at 4-bit or even lower precision—significantly reduces transfer volume, it typically results in sacrifice in precision and degradation of the final model accuracy. In this work, we introduce DUO, a distributed training framework designed to mitigate accuracy degradation incurred by gradient compression without involving additional overhead. DUO achieves this by inserting an additional high-precision gradient synchronization step into a previously computation-only phase, so that its communication is fully hidden by computation. We provide a comprehensive theoretical proof of convergence for DUO and validate its effectiveness through extensive pre-training experiments on GPT models. Our results indicate that DUO effectively restores accuracy when using 4-bit gradient compression, achieving performance comparable to uncompressed training. Remarkably, DUO maintains minimal accuracy degradation even under extreme compression scenarios, including 1-bit gradients or complete omission of the low-precision gradient communication step (0-bit transmission).

Scaling of deep neural networks, especially Transformers, is pivotal for their surging performance and has further led to the emergence of sophisticated reasoning capabilities in foundation models. Such scaling generally requires training large models from scratch with random initialization, failing to leverage the knowledge acquired by their smaller counterparts, which are already resource-intensive to obtain. To tackle this inefficiency, we present $\textbf{L}$ossl$\textbf{E}$ss $\textbf{MO}$del Expansio$\textbf{N}$ (LEMON), a recipe to initialize scaled models using the weights of their smaller but pre-trained counterparts. This is followed by model training with an optimized learning rate scheduler tailored explicitly for the scaled models, substantially reducing the training time compared to training from scratch. Notably, LEMON is versatile, ensuring compatibility with various network structures, including models like Vision Transformers and BERT. Our empirical results demonstrate that LEMON reduces computational costs by 56.7\% for Vision Transformers and 33.2\% for BERT when compared to training from scratch.

Recent years have witnessed a clear trend towards language models with an ever-increasing number of parameters, as well as the growing training overhead and memory usage. Distributed training, particularly through Sharded Data Parallelism (ShardedDP) which partitions optimizer states among workers, has emerged as a crucial technique to mitigate training time and memory usage. Yet, a major challenge in the scalability of ShardedDP is the intensive communication of weights and gradients. While compression techniques can alleviate this issue, they often result in worse accuracy. Driven by this limitation, we propose SDP4Bit (Toward 4Bit Communication Quantization in Sharded Data Parallelism for LLM Training), which effectively reduces the communication of weights and gradients to nearly 4 bits via two novel techniques: quantization on weight differences, and two-level gradient smooth quantization. Furthermore, SDP4Bit presents an algorithm-system co-design with runtime optimization to minimize the computation overhead of compression. Additional to the theoretical guarantees of convergence, we empirically evaluate the accuracy of SDP4Bit on the pre-training of GPT models with up to 6.7 billion parameters, and the results demonstrate a negligible impact on training loss. Furthermore, speed experiments show that SDP4Bit achieves up to 4.08× speedup in end-to-end throughput on a scale of 128 GPUs.

The attention-based Transformers have been increasingly applied to audio classification because of their global receptive field and ability to handle long-term dependency. However, the existing frameworks which are mainly extended from the Vision Transformers are not perfectly compatible with audio signals. In this paper, we introduce a Causal Audio Transformer (CAT) consisting of a Multi-Resolution Multi- Feature (MRMF) feature extraction with an acoustic attention block for more optimized audio modeling. In addition, we propose a causal module that alleviates over-fitting, helps with knowledge transfer and improves interpretability. CAT obtains higher or comparable state-of-the-art classification performance on ESC50, AudioSet and UrbanSound8K datasets, and can be easily generalized to other Transformer- based models.

In this paper, we propose Discrete Cosin TransFormer (DCFormer) that directly learn semantics from DCT-based frequency domain representation. We first show that transformer-based networks are able to learn semantics directly from frequency domain representation based on discrete cosine transform (DCT) without compromising the performance. To achieve the desired efficiency-effectiveness trade-off, we then leverage an input information compression on its frequency domain representation, which highlights the visually significant signals inspired by JPEG compression. We explore different frequency domain downsampling strategies and show that it is possible to preserve the semantic meaningful information by strategically dropping the high-frequency components. The proposed DCFormer is tested on various downstream tasks including image classification, object detection and instance segmentation, and achieves state-of-the-art comparable performance with less FLOPs, and outperforms the commonly used backbone (e.g. SWIN) at similar FLOPs. Our ablation results also show that the proposed method generalizes well on different transformer backbones.

We consider the problem of computing the$k$k-means centers for a large high-dimensional dataset in the context of edge-based machine learning, where data sources offload machine learning computation to nearby edge servers.$k$k-Means computation is fundamental to many data analytics, and the capability of computing provably accurate$k$k-means centers by leveraging the computation power of the edge servers, at a low communication and computation cost to the data sources, will greatly improve the performance of these analytics. We propose to let the data sources send small summaries, generated by joint dimensionality reduction (DR), cardinality reduction (CR), and quantization (QT), to support approximate$k$k-means computation at reduced complexity and communication cost. By analyzing the complexity, the communication cost, and the approximation error of$k$k-means algorithms based on carefully designed composition of DR/CR/QT methods, we show that: (i) it is possible to compute near-optimal$k$k-means centers at a near-linear complexity and a constant or logarithmic communication cost, (ii) the order of applying DR and CR significantly affects the complexity and the communication cost, and (iii) combining DR/CR methods with a properly configured quantizer can further reduce the communication cost without compromising the other performance metrics. Our theoretical analysis has been validated through experiments based on real datasets.

We consider the problem of computing the k-means centers for a large high-dimensional dataset in the context of edge-based machine learning, where data sources offload machine learning computation to nearby edge servers. k-Means computation is fundamental to many data analytics, and the capability of computing provably accurate k-means centers by leveraging the computation power of the edge servers, at a low communication and computation cost to the data sources, will greatly improve the performance of these analytics. We propose to let the data sources send small summaries, generated by joint dimensionality reduction (DR) and cardinality reduction (CR), to support approximate k-means computation at reduced complexity and communication cost. By analyzing the complexity, the communication cost, and the approximation error of k-means algorithms based on state-of-the-art DR/CR methods, we show that: (i) in the single-source case, it is possible to achieve a near-optimal approximation at a near-linear complexity and a constant communication cost, (ii) in the multiple-source case, it is possible to achieve similar performance at a logarithmic communication cost, and (iii) the order of applying DR and CR significantly affects the complexity and the communication cost. Our findings are validated through experiments based on real datasets.

Coreset, which is a summary of the original dataset in the form of a small weighted set in the same sample space, provides a promising approach to enable machine learning over distributed data. Although viewed as a proxy of the original dataset, each coreset is only designed to approximate the cost function of a specific machine learning problem, and thus different coresets are often required to solve different machine learning problems, increasing the communication overhead. We resolve this dilemma by developing robust coreset construction algorithms that can support a variety of machine learning problems. Motivated by empirical evidence that suitably-weighted k -clustering centers provide a robust coreset, we harden the observation by establishing theoretical conditions under which the coreset provides a guaranteed approximation for a broad range of machine learning problems, and developing both centralized and distributed algorithms to generate coresets satisfying the conditions. The robustness of the proposed algorithms is verified through extensive experiments on diverse datasets with respect to both supervised and unsupervised learning problems.

Motivated by the need of solving machine learning problems over distributed datasets, we explore the use of \emph{coreset} to reduce the communication overhead. Coreset is a summary of the original dataset in the form of a small weighted set in the same sample space. Compared to other data summaries, coreset has the advantage that it can be used as a proxy of the original dataset. However, existing coreset construction algorithms are each tailor-made for a specific machine learning problem. Thus, to solve different machine learning problems, one has to collect coresets of different types, defeating the purpose of saving communication overhead. We resolve this dilemma by developing robust coreset construction algorithms based on k-means/median clustering, that give a provably good approximation for a broad range of machine learning problems with sufficiently continuous cost functions. Through evaluations on diverse datasets and machine learning problems, we verify the robust performance of the proposed algorithms.