Zhipeng Deng

Ping Chen, Zhuohong Deng, Ping Li et al.

Training large language models (LLMs) is often constrained by GPU memory limitations. To alleviate memory pressure, activation recomputation and data compression have been proposed as two major strategies. However, both approaches have limitations: recomputation introduces significant training overhead, while compression can lead to accuracy degradation and computational inefficiency when applied naively. In this paper, we propose Adacc, the first adaptive memory optimization framework that unifies activation recomputation and data compression to improve training efficiency for LLMs while preserving model accuracy. Unlike existing methods that apply static, rule-based strategies or rely solely on one technique, Adacc makes fine-grained, tensor-level decisions, dynamically selecting between recomputation, retention, and compression based on tensor characteristics and runtime hardware constraints. Adacc tackles three key challenges: (1) it introduces layer-specific compression algorithms that mitigate accuracy loss by accounting for outliers in LLM activations; (2) it employs a MILP-based scheduling policy to globally optimize memory strategies across layers; and (3) it integrates an adaptive policy evolution mechanism to update strategies during training in response to changing data distributions. Experimental results show that Adacc improves training throughput by 1.01x to 1.37x compared to state-of-the-art frameworks, while maintaining accuracy comparable to the baseline.

Zhipeng Deng, Yuqiao Yang, Kenji Suzuki

Federated Learning (FL) enables multiple institutes to train models collaboratively without sharing private data. Current FL research focuses on communication efficiency, privacy protection, and personalization and assumes that the data of FL have already been ideally collected. In medical scenarios, however, data annotation demands both expertise and intensive labor, which is a critical problem in FL. Active learning (AL), has shown promising performance in reducing the number of data annotations in medical image analysis. We propose a federated AL (FedAL) framework in which AL is executed periodically and interactively under FL. We exploit a local model in each hospital and a global model acquired from FL to construct an ensemble. We use ensemble-entropy-based AL as an efficient data-annotation strategy in FL. Therefore, our FedAL framework can decrease the amount of annotated data and preserve patient privacy while maintaining the performance of FL. To our knowledge, this is the first FedAL framework applied to medical images. We validated our framework on real-world dermoscopic datasets. Using only 50% of samples, our framework was able to achieve state-of-the-art performance on a skin-lesion classification task. Our framework performed better than several state-of-the-art AL methods under FL and achieved comparable performance to full-data FL.