Steinar Laenen

I am a research intern at Five in the Oxford research group. At the moment I am working on few-shot learning together with Dr. Luca Bertinetto. The group has close collaborative ties to the Torr Vision group at Oxford University.

I received my MSc in Artificial Intelligence graduate (with distinction) from the University of Edinburgh. I wrote my thesis on spectral clustering of directed graphs using Hermitian adjacency matrix representations, supervised by Dr. He Sun. For the thesis I received the Joint AI MSc Dissertation Prize, awarded to 2 out of 200+ students.

I did my bachelors (summa cum laude) in mathematics and computer science at Amsterdam University College, where I wrote my thesis on trying to understand parallel pathways of the human visual cortex using deep neural networks, supervised by Dr. Steven Scholte.

Email  /  CV (on request)  /  LinkedIn  /  Google Scholar  /  GitHub


I've done/I am doing projects in machine learning, computer vision, spectral graph theory, virtual agents, virtual reality, NLP, and cognitive science. Here is a selection of some projects.


On Episodes, Prototypical Networks, and Few-Shot Learning.
Steinar Laenen, Luca Bertinetto.
Meta Learning Workshop at Neural Information Processing Systems , 2020. Contributed Talk. Code available soon.

Episodic learning is a popular practice among researchers and practitioners interested in few-shot learning. In this paper, we investigate the usefulness of episodic learning in Prototypical Networks and Matching Networks, two of the most popular algorithms making use of this practice. By drawing comparisons with Neighbourhood Component Analysis, we found that, for Prototypical and Matching Networks, it is detrimental to use the episodic learning strategy of separating training samples between support and query set, as it is a data-inefficient way to exploit training batches.

Higher-Order Spectral Clustering of Directed Graphs.
Steinar Laenen, He Sun.
Neural Information Processing Systems , 2020. Poster video.

We study directed graphs (digraphs) whose clusters exhibit "structural" information amongst each other. Based on the Hermitian matrix representation of digraphs, we present a nearly-linear time algorithm for digraph clustering, and further show that our proposed algorithm can be implemented in sublinear time under reasonable assumptions. The significance of our theoretical work is demonstrated by extensive experimental results on the UN Comtrade Dataset and synthetic block models.


Directed Graph Clustering Using Hermitian Laplacians.
Steinar Laenen, He Sun. Master's Thesis, 2019. Received Joint Artificial Intelligence MSc Dissertation Prize, awarded to 2 out of 200+ students. code

We investigated the use of Hermitian adjacency matrix representations for spectral clustering of directed graphs. We proved that assuming the graph is well-clustered, k-means clustering can find a good partitioning of a cyclic graph pattern using only 1 eigenvector. We further perform experiments on the UN Comtrade database and show that the Hermitian adjacency matrix representation outperforms other spectral clustering methods for directed graphs.


Reinforcement Learning for the football RoboCup 2D Half Field Offense (HFO) task.
Coursework Project for my Reinforcement Learning course at the University of Edinburgh.

For the coursework we implemented and optimized a wide set of single-agent and multi-agent reinforcment learning algorithms, including some of the canonical ones (Q-learning, SARSA, Value Iteration). We also implemented a state-of-the-art deep RL algorithm ( Asynchronous one-step Q-learning) on which I received a perfect score for implementation and performance.


Supervised and Reinforcement Learning for end-to-end Controllers in Autonomous Cars.
Edinburgh University Formula Student AI Research Team

While doing my AI MSc at the University of Edinburgh, I was part of the autonomous driving team at Edinburgh University Forumula Student , where I co-led the research team. We implemented a deep ConvNet to drive the car autonomously in ROS simulation using supervised learning, and we implemented an end-to-end controller that learns to drive a car from scratch using Deep Reinforcement Learning.


Organisor of a talk series on societal impact of AI development
We Need to Talk About AI - Talk Series

At Edinburgh University I was also heavily involved in initiating and organising a series of panel discussions titled: We Need to Talk About AI, where we invited insightful speakers to discuss some pressing societal issues around AI development. I was involved in three events which related to Data & Privacy, the Future of Work and Autonomous Weapons. Due to the succes of the talk series it is now officially funded and endorsed by the Edinburgh Futures Institute.


Thermal Mapping of Icelandic Geothermal Surface Manifestations with a Drone
Grimur Bjornsson, Gunnar Grimsson, Ari Sigurdsson, Valdimar S. Laenen
44th Workshop on Geothermal Reservoir Engineering Stanford University, 2019

Drone equipped with RGB and infrared cameras are used to create orthomosaic maps of areas that contain geothermal lineaments. Research was done at Warm Arctic ehf.

I developed an anomaly detection tool to automatically detect areas of geothermal importance using deep autoencoders and a one-class SVM.

Interpreting Social Commitment in a Simulated Theater
Daniel Veutgen, Marco Massetti, Joy Rossi, Leonardo Veroli, Athena Ásgeirsdóttir, Gudrun Baldursdóttir, Sigrún Gissurardóttir, Gísli Gudmundsson, Tinna Sigurdardóttir, Valdimar Laenen, Hannes Högni Vilhjálmsson
Proceedings of the 18th International Conference on Intelligent Virtual Agents, 2018.  

The research consisted of designing and programming automated intelligent social behaviour for virtual agents in Unity3D and C#. A cool thing that we added to the project was the ability to 'become' one of the characters using an Oculus Rift, and being able to interact with them.


Visual Pathways from the Perspective of Emotion and Physical Feature Detection in Multilabel Deep Neural Networks
Steinar Laenen, Steven Scholte
Bachelor's Thesis, 2018.

We use deep neural networks trained on two orthogonal targets to investigate whether seperation of tasks in the brain can be explained.

Based on Jon Barron's website