Zhengyang Hu

Zhengyang Hu

PhD Candidate, Department of ECE, HKU

Talks and presentations

Foundation-style Methods for Real-Time Statistical Dependency Measurement and Its Applications

May 20, 2026

Seminar, Department of Electrical and Electronic Engineering, The University of Hong Kong,

Invited departmental seminar at HKU ECE (20 May 2026, 11:00 AM – 12:00 PM).

The talk covered InfoNet (ICML 2024, Oral) and InfoAtlas (ICML 2026), focusing on pretrained tabular models for statistical dependence estimation. It also discussed online diagnostics based on statistical dependence among hidden representations to identify Transformer training anomalies, including overfitting, data anomalies and loss spikes, with validation on NanoGPT, ViT and Pythia.

InfoNet: Neural Estimation of Mutual Information without Test-Time Optimization

July 01, 2024

Conference Oral, International Conference on Machine Learning (ICML),

Oral presentation at ICML 2024 for our paper InfoNet: Neural Estimation of Mutual Information without Test-Time Optimization.

The talk introduced InfoNet, a pretrained tabular model for dependence estimation between random variables, achieving 1,000× faster estimation and higher accuracy on low-dimensional benchmarks. Downstream applications include feature selection, image segmentation and keyframe extraction.