IAPR Technical Committee 2 (TC2)

Structural and Syntactical Pattern Recognition

RESEARCH SPOTLIGHT

This month's research spotlight features Prof. Xin Ning, Professor at the Institute of Semiconductors, Chinese Academy of Sciences.

Introduce yourself.

I am Xin Ning, a Professor at the Institute of Semiconductors, Chinese Academy of Sciences, and Deputy Director of the Artificial Intelligence and High-Speed Circuit Laboratory. My research focuses on pattern recognition, machine learning, and computer vision, with particular interests in biomimetic image cognitive computing, visual representation learning, and 2D/3D visual understanding.

Over the years, my research has centered on how to represent and recognize complex visual patterns by modelling their intrinsic structures and variations. This has led to studies on discriminative feature learning, covering-based representation, neural models, point-cloud understanding, and learning with limited or evolving data. More recently, I have also been exploring multimodal learning, neural network optimization, neuromorphic vision, and AI for Science. Across these directions, a recurring objective is to develop recognition models that are more structurally expressive, adaptable, and capable of generalizing beyond conventional closed-set settings.

How did you become interested in pattern recognition, and how has your research developed?

My early research started from feature representation and face recognition. At that time, I became interested in how different observations of the same object or identity should be organized in a feature space. Many recognition methods focus on finding discriminative boundaries between classes. I began to think about the problem from another angle: before deciding where a boundary should be placed, can we first understand how a pattern itself is distributed in the feature space? This idea appeared in our early work on Biomimetic Uncorrelated Locality Discriminant Projection (BULDP), where we explored relationships among samples through principles such as homology continuity and heterogeneous similarity.

It later led us to think about recognition from the perspective of covering rather than only partitioning. If samples from the same pattern occupy a meaningful region in a high- dimensional space, a model can try to learn a geometric structure that covers this region. This became the basis for our work on the High-order Coverage Function Neural Network (HCFNN) and later the hyper-sausage coverage function neuron. Instead of only changing network architectures or loss functions, we revisited the neuron itself and asked whether a neuron could serve as a flexible geometric unit for representing a pattern distribution.

The hyper-sausage neuron was designed to model more flexible regions in feature space, rather than relying only on simple geometric forms such as hyperplanes or hyperspheres. I think the main value of this work is the idea behind it: a neuron can be viewed not only as a computational unit, but also as a learnable geometric representation of a pattern region. This work was later recognized with the 2023 Outstanding Paper Award from Pattern Recognition, which was very encouraging to us. It showed that revisiting fundamental issues such as pattern structure, geometry, and neuron modelling can still open interesting directions in the deep-learning era.

What are you working on now in pattern recognition?

One direction of my current research is to extend our earlier ideas from 2D images to 3D and multimodal visual understanding. Point clouds are particularly interesting because they are sparse, unordered, and strongly dependent on geometry. In our work, we have studied local geometric representation, point-cloud classification and segmentation, and more recently cross-modal knowledge transfer between images and 3D point clouds. For me, this is still the same fundamental problem: what structure best represents a pattern, and how can that structure be preserved while making the representation discriminative?

Another direction is few-shot class-incremental learning. Real recognition systems cannot assume that all classes and sufficient training data are available from the beginning. They need to learn new concepts from very few examples while retaining previously acquired knowledge. In our recent work, we explored brain-inspired fast- and slow-update mechanisms for this setting. I find this problem particularly interesting because it asks how a category can remain stable while continuously incorporating new variations and new knowledge. In a broader sense, this is closely related to the continuity of patterns that motivated some of my earlier research.

More recently, I have also become interested in neural network optimization and model adaptation. With pretrained and foundation models, scratch training and fine-tuning have become two very different learning scenarios. In our recent work, we revisited whether they should really be treated with the same optimization strategy.

Our group is also extending these ideas to AI for Science and neuromorphic vision. The applications may change, but I hope to keep one principle consistent: start from the underlying pattern and structure, rather than simply applying a larger model to a new task.

Any message to the readers, particularly early-career researchers?

Pattern recognition has changed dramatically, from handcrafted features and statistical learning to deep networks, transformers, and foundation models. But I do not think its fundamental questions have disappeared. What defines a pattern? What should remain invariant when its appearance changes? How should samples of the same category be organized? And how can old knowledge remain stable while new knowledge is continuously learned? These questions are still very relevant, even though we now have very different tools to study them.

For early-career researchers, I would encourage you not only to follow the latest architectures or benchmark results, but also to think carefully about the scientific question behind the method. A particular model may be replaced quickly, while a good research question can remain important for many years. I also think that structure deserves continued attention. Real-world data contain geometry, topology, temporal relations, semantic relations, and physical constraints. Understanding these structures can help us build recognition systems that are not only more accurate, but also more data-efficient, adaptable, and interpretable.

Page updated

Google Sites

Report abuse