Title:	Representation Learning Methods in Computer Vision
DNr:	Berzelius-2026-274
Project Type:	LiU Berzelius
Principal Investigator:	Atsuto Maki <atsuto@kth.se>
Affiliation:	Kungliga Tekniska högskolan
Duration:	2026-10-23 – 2027-05-01
Classification:	20208
Keywords:

Abstract

This project is a continuation of Berzelius-2026-75. As the group continues to expand, we are requesting an increased allocation compared with the previous allocation period. Representation learning is a key component in modern computer vision, enabling models to automatically learn rich, high-level features from raw data. In this project, we aim to advance representation learning techniques in computer vision in the following areas: 1. Football Tracking Data Understanding In this stage, we focus on context-aware representation learning for skeleton-based soccer motion, learning player motion representations based on their teammates, opponents, the ball, and environment. These representations will be evaluated on downstream tasks, such as action recognition and possession detection. 2. Deep Continual Learning with Imbalanced Data for Intelligent Systems Building on the robust representation learning frameworks established in the previous cycle, this project will now continue to focus on Deep Continual Learning with class-Imbalanced Data. The objective is to develop techniques that can learn from continuous streams of data without catastrophic forgetting, while specifically addressing class imbalance. The research will target computer vision benchmarks and possibly extend to NLP. 3. Steerable Latent World Models We plan to keep the current data and model checkpoints to maintain the option for continuation. 4. Scalable Online Task-Free Continual Learning for Real-Time Adaptation in High-Stakes Dynamic Environments This project develops a scalable Online Task-Free Continual Learning framework to enable AI models to learn incrementally from streaming data while mitigating catastrophic forgetting. The focus is on high-stakes, dynamic environments, particularly healthcare and robotics, where models must adapt in real-time without full retraining from scratch, under strict computational and efficiency constraints. We will explore novel architectures to balance stability and plasticity while minimizing compute and memory overhead, with evaluations on evolving datasets, healthcare benchmarks, and robotic perception tasks under real-world deployment conditions. 5. Task-Agnostic Deep Continual Learning with Class Imbalanced Data Unlike Project 2, this project focuses on task-agnostic continual learning under class imbalance without assuming explicit task identities or task boundaries. The aim is to identify key factors that enable models to learn new tasks while avoiding performance loss on previously learnt tasks, with focus on long-tailed distributions and keeping storage and computations at a minimum. The project will start with feasibility studies, where promising directions will then be analyzed further and more fully developed. The project will mainly focus on computer vision tasks, with the overarching goal to enable flexible, and resource efficient intelligence for e.g. robotic vision.