Architectures for Higher-Order and Robust Intelligence Computing based on Low-Dimensional Materials

This thesis addresses challenges in deep learning complexity through innovative solutions. We present a crossbar processing method using dual-gated memtransistors, showcasing their unique advantages, such as persistent programming and active control. This approach enables advanced inference architectures, reducing operating power significantly, particularly in emerging layers like hypernetworks. Additionally, we introduce a programmable Gaussian-like memory transistor (GMT) for probabilistic inference, demonstrating precise control over output parameters. The GMT's programmability, retention performance, and mechanical flexibility simplify circuit design and have been successfully applied in localization and obstacle avoidance tasks. The study also examines transformers' impact on long-range sequence processing, emphasizing computational challenges and contributing to ongoing research on hardware acceleration and probabilistic inference techniques for improved deep learning performance.