Skip to main content

Xingyang Ni: Neural networks can improve image embeddings' distinctiveness, privacy, and compressibility

Tampere University
LocationHervanta Campus, Tietotalo, auditorium TB109 (Korkeakoulunkatu 1, Tampere) and remote connection
20.5.2022 12.00–16.00
Entrance feeFree of charge
Bust shot of the researcher. Photo by Yurui Wang
The past decade has witnessed rapid advancement in the field of machine learning. Significant progress has been made in academic research, and practical applications are ubiquitous in everyday life. In his doctoral dissertation, MSc (Tech) Xingyang Ni studied the feature representations extracted by neural networks. He focused especially on making the image embeddings distinctive, privacy-preserving, and compressible.

Feature extraction signifies the process of converting data from one feature space to another. Regardless of the type of machine learning algorithm used, extracting useful features plays a vital role. In his doctoral dissertation, Xingyang Ni thoroughly analyzed three aspects of feature representations.

Firstly, features should be distinctive. This means that features of samples from distinct categories ought to differ significantly. It is essential for image retrieval systems in which searching is performed over image features based on similarities. Given a query sample, an algorithm finds the most similar samples in a gallery set.

Secondly, features should be privacy-preserving, meaning that it should not be possible to deduce sensitive information from the features. By broadly adopting Machine Learning as a Service (MLaaS), privacy-preserving features can be utilized to prevent privacy violations in case the server is compromised. It is especially relevant for a model deployed in production as deep features would be transferred and stored.

Thirdly, features should be compressible. In other words, compact features are preferable as they require less storage space. Obtaining compressible features is of great importance in data compression. Note that a trade-off between fidelity and compression ratio has to be made.

"A diverse collection of studies has been included, covering areas such as person re-identification, vehicle attribute recognition, neural image compression, clustering, and unsupervised anomaly detection. Research articles included in my dissertation reveal different approaches to improving image embeddings learned by neural networks. This topic remains a fundamental challenge in Machine Learning, and further research is needed to gain a deeper understanding," says Ni.

The doctoral dissertation of MSc (Tech) Xingyang Ni in the field of information technology titled Towards Better Image Embeddings Using Neural Networks will be publicly examined in the Faculty of Information Technology and Communication Sciences at Tampere University at 12 o'clock on 20 May 2022. The venue is auditorium TB109 in Tietotalo (Korkeakoulunkatu 1, Tampere). The Opponent will be Professor Guoying Zhao from University of Oulu. The Custos will be Associate Professor Esa Rahtu from Tampere University. The dissertation is co-supervised by Dr. Heikki Huttunen from Visy Oy.

The event can be followed via remote connection.

The dissertation is available online at

Photo: Yurui Wang