Baidu Research Institute 2023 Paper Sharing Session – Presented Papers Overview

Event Time: December 19‑20, 2023, 14:00

Event: Baidu Research Institute 2023 Paper Sharing Session

The session features seven speakers from Baidu Research Institute’s Business Intelligence Lab, Cognitive Computing Lab, and Big Data Lab. They will present eight papers published at top conferences such as KDD, NeurIPS, ICCV, and others, covering topics like search ranking, vehicle‑infrastructure coordination, deep learning interpretability, and 3D scene reconstruction.

Report 1: S2phere: Semi‑Supervised Pre‑training for Web Search over Heterogeneous Learning to Rank Data

Speaker: Li Yuncheng (Baidu Big Data Lab, Ph.D. candidate at Shanghai Jiao Tong University)

Abstract: The paper addresses the scarcity of high‑quality labeled data for commercial search ranking. It proposes a semi‑supervised deep ranking algorithm that leverages heterogeneous data, and the work was published at KDD 2023.

Report 2: Hierarchical Reinforcement Learning for Dynamic Autonomous Vehicle Navigation at Intelligent Intersections

Speaker: Zhang Le (Baidu Business Intelligence Lab)

Abstract: Introduces the NavTL framework, a reinforcement‑learning‑based method that jointly controls traffic signals and plans routes for autonomous vehicles, reducing intersection congestion and improving overall travel efficiency. Experiments on real‑world and synthetic datasets demonstrate its effectiveness.

Report 3: Multi‑Temporal Relationship Inference in Urban Areas

Speaker: Li Shuangli (Baidu Business Intelligence Lab, Ph.D. student at University of Science and Technology of China)

Abstract: Proposes SEENet, a spatial‑evolution graph neural network that models dynamic relationships among urban locations, integrating temporal and geographic influences. Experiments on four city datasets show superior performance in dynamic relationship mining.

Report 4: Spatial Heterophily Aware Graph Neural Networks

Speaker: Xiao Congxi (Baidu Business Intelligence Lab, Ph.D. student at University of Science and Technology of China)

Abstract: Highlights the spatial heterophily present in urban graphs and introduces a graph neural network that groups neighbors by spatial relations to mitigate distribution disparities, improving performance on heterogeneous urban data.

Report 5: M4: A Unified XAI Benchmark for Faithfulness Evaluation of Feature Attribution Methods across Metrics, Modalities and Models

Speaker: Chen Jiamin (Baidu Big Data Lab)

Abstract: Presents the M4 benchmark for evaluating the faithfulness of feature‑attribution methods across different metrics, modalities, and model families, and provides extensive experimental comparisons.

Report 6: Differentiable Neuro‑Symbolic Reasoning on Large‑Scale Knowledge Graphs

Speaker: Fang Huang (Baidu Cognitive Computing Lab)

Abstract: Introduces DiffLogic, a differentiable neuro‑symbolic reasoning framework built on Probabilistic Soft Logic, enabling end‑to‑end optimization of KG embeddings and logical rule weights, outperforming baselines on standard benchmarks.

Report 7: On the Overlooked Structure of Stochastic Gradients

Speaker: Xie Zeke (Baidu Cognitive Computing Lab)

Abstract: Statistical analysis reveals a power‑law structure in stochastic gradients, explaining low‑dimensional dynamics in deep learning and drawing parallels with protein dynamics.

Report 8: S3IM: Stochastic Structural Similarity and Its Unreasonable Effectiveness for Neural Fields

Speaker: Xie Zeke (Baidu Cognitive Computing Lab)

Abstract: Proposes S3IM, a stochastic structural loss that captures non‑local information in neural fields such as NeRF, achieving significant reconstruction quality improvements with negligible computational overhead.