Yang Luo

Yang Luo, Xiaozhe Ren, Zangwei Zheng et al.

Adaptive gradient methods, such as Adam and LAMB, have demonstrated excellent performance in the training of large language models. Nevertheless, the need for adaptivity requires maintaining second-moment estimates of the per-parameter gradients, which entails a high cost of extra memory overheads. To solve this problem, several memory-efficient optimizers (e.g., Adafactor) have been proposed to obtain a drastic reduction in auxiliary memory usage, but with a performance penalty. In this paper, we first study a confidence-guided strategy to reduce the instability of existing memory efficient optimizers. Based on this strategy, we propose CAME to simultaneously achieve two goals: fast convergence as in traditional adaptive methods, and low memory usage as in memory-efficient methods. Extensive experiments demonstrate the training stability and superior performance of CAME across various NLP tasks such as BERT and GPT-2 training. Notably, for BERT pre-training on the large batch size of 32,768, our proposed optimizer attains faster convergence and higher accuracy compared with the Adam optimizer. The implementation of CAME is publicly available.

Yang Luo, Zangwei Zheng, Ziheng Qin et al.

Large-batch training has become a cornerstone in accelerating the training of deep neural networks, yet it poses challenges in optimization and generalization. Existing optimizers like AdamW present performance degradation during language models' large-batch training, due to the information bottleneck in attention layers caused by the sharp increase of max attention logit. While the LAMB optimizer partially addresses this issue, some attention layers still face this issue. The reason is that $l_2$-norm-based trust ratios in LAMB are less effective in directly influencing the max value of query/key weights. Furthermore, the weight-wise trust ratio in LAMB is error-prone as it overlooks relationships of weight values within rows or columns. Building on these observations, we propose a novel optimizer, MERIT, which leverages the max-norm to calculate the trust ratio to constrain the max attention logit more effectively. Moreover, we further construct element-wise trust ratios to provide more robust update scaling by focusing on local weight structures. Extensive experiments of large-batch training across various sizes of GPT-2 models demonstrate the superior performance of MERIT. Notably, during the training of GPT-2 Medium, MERIT enables a 6k batch size without any performance degradation compared to the standard batch size (480) with 48B training tokens. This work highlights the importance of considering the max attention logit and finer-granularity trust ratio in large-batch training. It successfully improves the training stability and paves the way for larger batch usage, enabling faster development and iteration of large language models. Code is available at https://github.com/NUS-HPC-AI-Lab/MERIT.

Yang Luo, Zangwei Zheng, Zirui Zhu et al.

The increase in parameter size of multimodal large language models (MLLMs) introduces significant capabilities, particularly in-context learning, where MLLMs enhance task performance without updating pre-trained parameters. This effectiveness, however, hinges on the appropriate selection of in-context examples, a process that is currently biased towards visual data, overlooking textual information. Furthermore, the area of supervised retrievers for MLLMs, crucial for optimal in-context example selection, continues to be uninvestigated. Our study offers an in-depth evaluation of the impact of textual information on the unsupervised selection of in-context examples in multimodal contexts, uncovering a notable sensitivity of retriever performance to the employed modalities. Responding to this, we introduce a novel supervised MLLM-retriever MSIER that employs a neural network to select examples that enhance multimodal in-context learning efficiency. This approach is validated through extensive testing across three distinct tasks, demonstrating the method's effectiveness. Additionally, we investigate the influence of modalities on our supervised retrieval method's training and pinpoint factors contributing to our model's success. This exploration paves the way for future advancements, highlighting the potential for refined in-context learning in MLLMs through the strategic use of multimodal data.

Yang Luo, Qingni Shen, Zhonghai Wu

Access control is an important component for web services such as a cloud. Current clouds tend to design the access control mechanism together with the policy language on their own. It leads to two issues: (i) a cloud user has to learn different policy languages to use multiple clouds, and (ii) a cloud service provider has to customize an authorization mechanism based on its business requirement, which brings high development cost. In this work, a new access control policy language called PERM modeling language (PML) is proposed to express various access control models such as access control list (ACL), role-based access control (RBAC) and attribute-based access control (ABAC), etc. PML's enforcement mechanism is designed in an interpreter-on-interpreter manner, which not only secures the authorization code with sandboxing, but also extends PML to all programming languages that support Lua. PML is already adopted by real-world projects such as Intel's RMD, VMware's Dispatch, Orange's Gobis and so on, which proves PML's usability. The performance evaluation on OpenStack, CloudStack and Amazon Web Services (AWS) shows PML's enforcement overhead per request is under 5.9us.

Yang Luo, Qixun Zhang, Qingni Shen et al.

With the expansion of the market share occupied by the Android platform, security issues (especially application security) have become attention focus of researchers. In fact, the existing methods lack the capabilities to manage application permissions without root privilege. This study proposes a dynamic management mechanism of Android application permissions based on security policies. The paper first describes the permissions by security policies, then implementes permission checking code and request evaluation algorithm in Android framework layer. Experimental results indicate that the presented approach succeeds in permission management of Android applications, and its system overhead is low, which makes it an effective method for Android permission management.