Lin Yudeng, Gao Bin, Tang Jianshi, Zhang Qingtian, Qian He, Wu Huaqiang
School of Integrated Circuits, Beijing National Research Center for Information Science and Technology, Tsinghua University, Beijing, China.
Nat Comput Sci. 2025 Jan;5(1):27-36. doi: 10.1038/s43588-024-00744-y. Epub 2024 Dec 23.
Labeling data is a time-consuming, labor-intensive and costly procedure for many artificial intelligence tasks. Deep Bayesian active learning (DBAL) boosts labeling efficiency exponentially, substantially reducing costs. However, DBAL demands high-bandwidth data transfer and probabilistic computing, posing great challenges for conventional deterministic hardware. Here we propose a memristor stochastic gradient Langevin dynamics in situ learning method that uses the stochastic of memristor modulation to learn efficiency, enabling DBAL within the computation-in-memory (CIM) framework. To prove the feasibility and effectiveness of the proposed method, we implemented in-memory DBAL on a memristor-based stochastic CIM system and successfully demonstrated a robot's skill learning task. The inherent stochastic characteristics of memristors allow a four-layer memristor Bayesian deep neural network to efficiently identify and learn from uncertain samples. Compared with cutting-edge conventional complementary metal-oxide-semiconductor-based hardware implementation, the stochastic CIM system achieves a remarkable 44% boost in speed and could conserve 153 times more energy.
对于许多人工智能任务而言，标注数据是一项耗时、费力且成本高昂的过程。深度贝叶斯主动学习（DBAL）能大幅提高标注效率，显著降低成本。然而，DBAL需要高带宽数据传输和概率计算，这给传统确定性硬件带来了巨大挑战。在此，我们提出一种忆阻器随机梯度朗之万动力学原位学习方法，该方法利用忆阻器调制的随机性来提高学习效率，从而在内存计算（CIM）框架内实现DBAL。为证明所提方法的可行性和有效性，我们在基于忆阻器的随机CIM系统上实现了内存中的DBAL，并成功展示了机器人的技能学习任务。忆阻器固有的随机特性使四层忆阻器贝叶斯深度神经网络能够有效地从不确定样本中识别并学习。与基于传统前沿互补金属氧化物半导体的硬件实现相比，随机CIM系统的速度显著提高了44%，且能节省153倍以上的能量。
