Export
He Xiao
· 3 previous affiliations
- University of Hong KongSince 2024
- PhD student · 2024–2028
- Current
- Southwest University2016–2024
- Researcher · 2021–2024
- Undergrad student · 2016–2020
- The University of Western Australia2019–2021
- MS student · 2019–2021
Research Overview
He Xiao focuses on the intersection of advanced artificial intelligence algorithms and high-performance hardware architectures, with a specific emphasis on in-memory computing (CIM) and ultra-low-bit quantization for large language models (LLMs). Their work addresses the computational bottlenecks of deploying foundation models on edge devices by developing hardware-software co-optimization strategies, particularly through memristor-based reconfigurable circuits and post-training quantization frameworks.
Research Trajectory
The researcher’s early work, primarily conducted during their Master’s and initial research assistant roles between 2022 and 2023, centered on the development of memristor-based neuromorphic circuits. During this period, the focus was on implementing fundamental neural structures—such as attention mechanisms and Transformers—directly into hardware to bypass the von Neumann bottleneck. This phase was characterized by an exploration of how non-volatile memory elements could be used to perform Multiplication Free Inference with Ternary Neural Networks and other low-precision operations. Key contributions included the design of light-weight Transformer circuits specifically optimized for speech recognition tasks and the development of multimodal in-memory computing frameworks for edge AI.
Starting in 2024, coinciding with their doctoral studies at The University of Hong Kong, the researcher’s focus shifted toward the challenges of scaling these efficiency gains to large language models. As LLMs grew in parameter count, the limitations of traditional quantization became apparent, particularly the "performance cliff" encountered when compressing models below 4 bits. This led to a research program dedicated to High Accuracy Model Compression Without Retraining, utilizing post-training quantization (PTQ) techniques. The researcher began investigating structured ternary representations and bit-plane decomposition to maintain model expressiveness at extremely low bit-widths.
Currently, the work has evolved into a comprehensive hardware-software co-design approach. This involves not only the algorithmic side of quantization but also the architectural design of specialized AI accelerators. Recent projects include the development of token-aware in-memory attention architectures and the exploration of Optimal Reconstruction for Low-Bit Neural Networks. This trajectory reflects a move from component-level memristive circuit design to system-level optimization for the most demanding modern AI workloads.
Current Directions
The researcher is currently pursuing three primary lines of inquiry: ultra-low-bit quantization frameworks, hardware-aware attention mechanisms, and ASIC design for edge deployment.
In the domain of quantization, the focus is on overcoming the representational limits of binary and ternary weights. By decomposing weight matrices into structured "trit-planes," the researcher aims to achieve the efficiency of multiplication-free inference without the significant accuracy loss typical of ultra-low-bit PTQ. This involves developing new mathematical frameworks for weight reconstruction that can operate without the need for extensive fine-tuning or retraining, which is often computationally prohibitive for models with billions of parameters.
Simultaneously, the researcher is investigating "Token-Aware" architectures. Recognizing that not all tokens in a sequence require the same level of computational precision or memory bandwidth, this work seeks to design in-memory attention mechanisms that can dynamically allocate resources. This hardware-software co-optimization is intended to reduce the energy footprint of the self-attention mechanism, which remains a primary bottleneck in LLM inference.
Finally, the researcher is engaged in practical ASIC (Application-Specific Integrated Circuit) research. This includes participating in tape-out projects to validate theoretical CIM architectures in physical silicon. The goal is to bridge the gap between high-level algorithmic quantization and the physical constraints of edge hardware, such as power consumption and silicon area.
Representative Work
The following papers represent the evolution of the researcher's contributions to the field, moving from memristive circuit design to advanced LLM quantization.
In PTQTP: Post-Training Quantization to Trit-Planes for Large Language Models, the researcher serves as the lead author, introducing a framework that decomposes weight matrices into structured ternary ({-1, 0, 1}) planes. This work addresses the trade-off between computational efficiency and model expressiveness in ultra-low-bit settings. By utilizing a 2×1.58-bit representation, the framework enables multiplication-free inference while mitigating the performance collapse often seen in 2-bit quantization. This represents a significant step in making LLMs viable for resource-constrained edge devices.
The researcher also participated in the development of BPDQ: Bit-Plane Decomposition Quantization on a Variable Grid for Large Language Models. As a middle author, they contributed to the exploration of bit-plane decomposition on variable grids. This research targets the "holy grail" of 2-bit and 3-bit quantization, seeking to maintain reasoning capabilities in massive models without the need for the high-precision floating-point arithmetic that typically dominates deep learning inference.
Earlier first-author work, such as "Brain-Inspired Recognition System Based on Multimodal In-Memory Computing Framework for Edge AI" (published in IEEE TCAS-I, 2024), established the researcher's foundation in CIM. This paper detailed a multimodal framework designed to handle diverse data types within a unified in-memory computing architecture, emphasizing the energy efficiency required for edge-based recognition systems.
In "Efficient Low-bit Neural Network with Memristor-Based Reconfigurable Circuits" (IEEE TCAS-II, 2024), the researcher explored the hardware implementation of low-bit networks. This work focused on the design of reconfigurable circuits that could adapt to different quantization levels, providing a hardware-level solution to the problem of balancing precision and power consumption.
The researcher's focus on Transformer architectures is evidenced in "Memristor-Based Light-Weight Transformer Circuit Implementation for Speech Recognition" (IEEE JETCAS, 2023). As the first author, they proposed a circuit-level implementation of the Transformer's self-attention mechanism using memristor crossbars. This work was one of the earlier attempts to map the complex dependencies of Transformers onto non-von Neumann hardware specifically for speech processing tasks.
Finally, the researcher has led the development of "TIA: Token-Aware In-Memory Attention Architecture for Large Language Models" (currently under review). This work represents the culmination of their current research interests, proposing a specialized architecture that optimizes the attention mechanism by considering the varying importance of different tokens during inference. This approach aims to significantly reduce the memory and computational overhead of LLMs in hardware.
While the broader bibliographic record includes early entries in fields such as meteorology and materials science (dating back to 2000), these appear to be associated with different individuals. The researcher's own academic trajectory is firmly rooted in the modern evolution of efficient AI hardware and LLM optimization.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake