The Institute of Artificial Intelligence of Peking University is one of the top artificial intelligence research institutions in China, and its researchers include many top scientists and engineers.

The institute mainly conducts research from the following three levels

· Supportive and Cross-disciplinary Level: Comprising support (intelligent software systems, intelligent neuromorphic chips, visual perception) and cross-disciplinary (mathematical and physical foundations, artificial intelligence governance, computational social science, intelligent healthcare, smart health).
· Key Domain Level: Including computer vision, natural language processing, computational cognition and common-sense reasoning, multi-agent systems, robotics research, and machine learning.
· General Platform Level: Building a general artificial intelligence system platform and a large-scale task testing platform.

Studies


ACCURATE YET EFFICIENT STOCHASTIC

COMPUTING NEURAL ACCELERATION WITH HIGH PRECISION RESIDUAL FUSION

A neural network accelerator based on stochastic computing is proposed. Through collaborative design and optimization of the neural network and accelerator chip, high-precision residual connections are fused into low-precision stochastic computing convolution operations. This effectively enhances the inference accuracy (by 9.43%) and almost incurs negligible additional hardware overhead (only 1.3%).

NEURALKOOPMAN POOLING: CONTROL-INSPIRED TEMPORAL DYNAMICS ENCODING FOR SKELETON-BASED ACTION RECOGNITION

Addressing the problem that existing pooling layers in deep neural networks fail to capture high-order dynamic information in sequential data, this paper proposes a plug-and-play parameterized pooling module based on the Koopman operator. The nonlinear system is linearized, representing complex systems using a linear evolution matrix. A method of eigenvalue regularization is introduced to impose stability constraints on the learned linear systems.

SAFE MULTI-AGENT REINFORCEMENT LEARNING FOR MULTI-ROBOT CONTROL

A challenging problem in robotics is how to control multiple robots cooperatively and safely in real-world applications.In this study，researchers investigate safe MARL for multi-robot control on cooperative tasks, in which each individual robot has to not only meet its own safety constraints while maximising their reward, but also consider those of others to guarantee safe team behaviours.

EMERGENT GRAPHICAL CONVENTIONS IN A

VISUAL COMMUNICATION GAME

Using the "Draw and Guess" game, a novel graphic symbol system was emerged and evolved computationally for the first time. Additionally, three attributes of graphic symbols were proposed: iconicity, symbolism, and semantics. This work offers a fresh computational framework and perspective for researching the origin and evolution of human language and writing.

HUMANISE: LANGUAGE-CONDITIONED HUMAN MOTION GENERATION IN 3D SCENES

This study proposed a large-scale HSI dataset with rich semantic annotations, named HUMANISE. This endeavor also initiated a novel task: generating human motion sequences in three-dimensional scenes under linguistic constraints. The paper further devises a motion generation model with scene and language constraints, capable of generating diverse and semantically coherent indoor human motions.

A DISTRIBUTED NANOCLUSTER BASED MULTI-AGENTEVOLUTIONARY NETWORK

Starting from the growth kinetics of conductive filaments regulated by electric fields, this study utilizes the inherent physical evolution patterns of memristive devices to construct, for the first time, a microscale computational system capable of highly mapping collective intelligent behaviors of multi-agent self-organized evolution.