Lintao Zhang

Zhiqi Lin, Youshan Miao, Guodong Liu et al. · microsoft-research

With the growing model size, deep neural networks (DNN) are increasingly trained over massive GPU accelerators, which demands a proper parallelization plan that transforms a DNN model into fine-grained tasks and then schedules them to GPUs for execution. Due to the large search space, the contemporary parallelization plan generators often rely on empirical rules that couple transformation and scheduling, and fall short in exploring more flexible schedules that yield better memory usage and compute efficiency. This tension can be exacerbated by the emerging models with increasing complexity in their structure and model size. SuperScaler is a system that facilitates the design and generation of highly flexible parallelization plans. It formulates the plan design and generation into three sequential phases explicitly: model transformation, space-time scheduling, and data dependency preserving. Such a principled approach decouples multiple seemingly intertwined factors and enables the composition of highly flexible parallelization plans. As a result, SuperScaler can not only generate empirical parallelization plans, but also construct new plans that achieve up to 3.5X speedup compared to state-of-the-art solutions like DeepSpeed, Megatron and Alpa, for emerging DNN models like Swin-Transformer and AlphaFold2, as well as well-optimized models like GPT-3.

Hong Wang, Yuanzhi Zhou, Chi Zhang et al.

This paper introduces XFL, an industrial-grade federated learning project. XFL supports training AI models collaboratively on multiple devices, while utilizes homomorphic encryption, differential privacy, secure multi-party computation and other security technologies ensuring no leakage of data. XFL provides an abundant algorithms library, integrating a large number of pre-built, secure and outstanding federated learning algorithms, covering both the horizontally and vertically federated learning scenarios. Numerical experiments have shown the prominent performace of these algorithms. XFL builds a concise configuration interfaces with presettings for all federation algorithms, and supports the rapid deployment via docker containers.Therefore, we believe XFL is the most user-friendly and easy-to-develop federated learning framework. XFL is open-sourced, and both the code and documents are available at https://github.com/paritybit-ai/XFL.

Lintao Zhang, Jinjian Wu, Lihong Wang et al.

Brain structural MRI has been widely used to assess the future progression of cognitive impairment (CI). Previous learning-based studies usually suffer from the issue of small-sized labeled training data, while there exist a huge amount of structural MRIs in large-scale public databases. Intuitively, brain anatomical structures derived from these public MRIs (even without task-specific label information) can be used to boost CI progression trajectory prediction. However, previous studies seldom take advantage of such brain anatomy prior. To this end, this paper proposes a brain anatomy prior modeling (BAPM) framework to forecast the clinical progression of cognitive impairment with small-sized target MRIs by exploring anatomical brain structures. Specifically, the BAPM consists of a pretext model and a downstream model, with a shared brain anatomy-guided encoder to model brain anatomy prior explicitly. Besides the encoder, the pretext model also contains two decoders for two auxiliary tasks (i.e., MRI reconstruction and brain tissue segmentation), while the downstream model relies on a predictor for classification. The brain anatomy-guided encoder is pre-trained with the pretext model on 9,344 auxiliary MRIs without diagnostic labels for anatomy prior modeling. With this encoder frozen, the downstream model is then fine-tuned on limited target MRIs for prediction. We validate the BAPM on two CI-related studies with T1-weighted MRIs from 448 subjects. Experimental results suggest the effectiveness of BAPM in (1) four CI progression prediction tasks, (2) MR image reconstruction, and (3) brain tissue segmentation, compared with several state-of-the-art methods.

Mengqi Wu, Lintao Zhang, Pew-Thian Yap et al.

Brain magnetic resonance imaging (MRI) has been extensively employed across clinical and research fields, but often exhibits sensitivity to site effects arising from non-biological variations such as differences in field strength and scanner vendors. Numerous retrospective MRI harmonization techniques have demonstrated encouraging outcomes in reducing the site effects at the image level. However, existing methods generally suffer from high computational requirements and limited generalizability, restricting their applicability to unseen MRIs. In this paper, we design a novel disentangled latent energy-based style translation (DLEST) framework for unpaired image-level MRI harmonization, consisting of (a) site-invariant image generation (SIG), (b) site-specific style translation (SST), and (c) site-specific MRI synthesis (SMS). Specifically, the SIG employs a latent autoencoder to encode MRIs into a low-dimensional latent space and reconstruct MRIs from latent codes. The SST utilizes an energy-based model to comprehend the global latent distribution of a target domain and translate source latent codes toward the target domain, while SMS enables MRI synthesis with a target-specific style. By disentangling image generation and style translation in latent space, the DLEST can achieve efficient style translation. Our model was trained on T1-weighted MRIs from a public dataset (with 3,984 subjects across 58 acquisition sites/settings) and validated on an independent dataset (with 9 traveling subjects scanned in 11 sites/settings) in four tasks: histogram and feature visualization, site classification, brain tissue segmentation, and site-specific structural MRI synthesis. Qualitative and quantitative results demonstrate the superiority of our method over several state-of-the-arts.

Kai-Fung Chu, Lintao Zhang

Many application scenarios call for training a machine learning model among multiple participants. Federated learning (FL) was proposed to enable joint training of a deep learning model using the local data in each party without revealing the data to others. Among various types of FL methods, vertical FL is a category to handle data sources with the same ID space and different feature spaces. However, existing vertical FL methods suffer from limitations such as restrictive neural network structure, slow training speed, and often lack the ability to take advantage of data with unmatched IDs. In this work, we propose an FL method called self-taught federated learning to address the aforementioned issues, which uses unsupervised feature extraction techniques for distributed supervised deep learning tasks. In this method, only latent variables are transmitted to other parties for model training, while privacy is preserved by storing the data and parameters of activations, weights, and biases locally. Extensive experiments are performed to evaluate and demonstrate the validity and efficiency of the proposed method.

Xiaotian Gao, Cui Wei, Lintao Zhang et al.

Training and inference efficiency of deep neural networks highly rely on the performance of tensor operators on hardware platforms. Manually optimizing tensor operators has limitations in terms of supporting new operators or hardware platforms. Therefore, automatically optimizing device code configurations of tensor operators is getting increasingly attractive. However, current methods for tensor operator optimization usually suffer from poor sample-efficiency due to the combinatorial search space. In this work, we propose a novel evolutionary method, OpEvo, which efficiently explores the search spaces of tensor operators by introducing a topology-aware mutation operation based on q-random walk to leverage the topological structures over the search spaces. Our comprehensive experiment results show that compared with state-of-the-art (SOTA) methods OpEvo can find the best configuration with the lowest variance and least efforts in the number of trials and wall-clock time. All code of this work is available online.

Jilong Xue, Youshan Miao, Cheng Chen et al.

Deep learning emerges as an important new resource-intensive workload and has been successfully applied in computer vision, speech, natural language processing, and so on. Distributed deep learning is becoming a necessity to cope with growing data and model sizes. Its computation is typically characterized by a simple tensor data abstraction to model multi-dimensional matrices, a data-flow graph to model computation, and iterative executions with relatively frequent synchronizations, thereby making it substantially different from Map/Reduce style distributed big data computation. RPC, commonly used as the communication primitive, has been adopted by popular deep learning frameworks such as TensorFlow, which uses gRPC. We show that RPC is sub-optimal for distributed deep learning computation, especially on an RDMA-capable network. The tensor abstraction and data-flow graph, coupled with an RDMA network, offers the opportunity to reduce the unnecessary overhead (e.g., memory copy) without sacrificing programmability and generality. In particular, from a data access point of view, a remote machine is abstracted just as a "device" on an RDMA channel, with a simple memory interface for allocating, reading, and writing memory regions. Our graph analyzer looks at both the data flow graph and the tensors to optimize memory allocation and remote data access using this interface. The result is up to 25 times speedup in representative deep learning benchmarks against the standard gRPC in TensorFlow and up to 169% improvement even against an RPC implementation optimized for RDMA, leading to faster convergence in the training process.

Zichuan Lin, Tianqi Zhao, Guangwen Yang et al.

Reinforcement learning (RL) algorithms have made huge progress in recent years by leveraging the power of deep neural networks (DNN). Despite the success, deep RL algorithms are known to be sample inefficient, often requiring many rounds of interaction with the environments to obtain satisfactory performance. Recently, episodic memory based RL has attracted attention due to its ability to latch on good actions quickly. In this paper, we present a simple yet effective biologically inspired RL algorithm called Episodic Memory Deep Q-Networks (EMDQN), which leverages episodic memory to supervise an agent during training. Experiments show that our proposed method can lead to better sample efficiency and is more likely to find good policies. It only requires 1/5 of the interactions of DQN to achieve many state-of-the-art performances on Atari games, significantly outperforming regular DQN and other episodic memory based RL algorithms.