Zhengyang Yu

I'm a fourth-year PhD student at Frankfurt Institute for Advanced Studies (FIAS), supervised by Prof. Dr. Jochen Triesch. I also collaborate closely with Prof. Chen Yu, and members of Developing Intelligence Lab. Before that, I received my Master’s degree from Xidian University, where I focused on signal processing and pattern recognition. In addition, I completed a short-term internship at Thoughtworks.

Email  /  CV  /  Scholar  /  Github

Research

My research interests include self-supervised learning, representation learning, and gaze behavior. I am currently developing more robust self-supervised models inspired by infants’ visual behaviors. I am also interested in multimodal infant data, such as video, audio, EEG, and ECG signals.

Cortical magnification creates a trade-off between enlarging objects and compressing the background, which supports self-supervised object learning.

We constructed a series of egocentric toddler gaze datasets and demonstrated, through representation learning, that toddlers' gaze behavior supports self-supervised object learning.

A novel foveation-inspired positional encoding for ViTs that reduces the number of learnable parameters.

We reveal distinct naturalistic saccade patterns and explain them with a cortical magnification–based model.

A biologically inspired contrastive learning framework that leverages sequential views instead of arbitrary augmentations to achieve near-supervised object recognition performance.

A novel metric for evaluating CNN performance in radar time–frequency spectrogram recognition.

A lightweight incremental ensemble learning algorithm that integrates new categories without retraining, significantly reducing training costs in noisy emitter classification.

A contour-extraction-based CNN method for radar intra-pulse modulation recognition, which simplifies the network structure and improves accuracy.

An SSA-based life signal detection method using 0.33 THz radar, which outperforms EMD under low SNR and shows promise for remote patient monitoring.

Miscellanea

Design and source code from Jon Barron's website.