Supervision: Andreas Loukas

Project type: Master thesis

Finished

I am looking for motivated students that wish to work on the development of deep learning algorithms for graph data.

GRAPH NEURAL NETWORKS

The capacity of machines to recognize patterns and make predictions has begun to rival (or even exceed) that of humans. Indeed, deep learning has been very successful in teaching machines how to learn from data with regular structure, such as images, audio, and text. Still, much of the data generated by our society exhibit graph structure. Take the human brain as an example---rather than looking at the distribution of neurons in space, scientists currently believe that the secret in decoding how the brain functions is by looking at the graph of its (functional and structural) organization. Similarly, the diffusion of information in a social network and the evolution of traffic are better understood, respectively, by how people and roads interconnect.

Realizing that for graph data standard deep learning falls short, researchers have attempted to replicate the success of computer vision by generalizing concepts from the pixel-grid to graphs. Building on powerful ideas from graph signal processing and manifold learning, we have witnessed the emergence of graph neural networks (GNN) --- neural network architectures that exploit the inherent properties of a graph to shape and regularize the learning process.

WHAT SHOULD BE EXPECTED

The student will work on an open research problem. He/she will interact frequently (on a bi-weekly basis) with the supervisor, receiving ample direction and advice. The work will resolve around the following tasks:

  • data gathering and pre-processing (e.g., social network data, bio-informatics data)

  • getting up to speed with the literature (paper reading)

  • development of novel deep learning algorithms (most likely in pytorch geometric)

  • thorough testing and comparison with SoA

  • Paper writing

The balance of the above will be tailored to the strengths and preferences of the student.

IDEAL CANDIDATE PROFILE

The ideal candidate will have experience with:

  • graph theory

  • python

  • machine learning (especially deep learning)

In addition, he/she will be a fast learner and driven, not afraid of hard work, and with a strong desire to do cutting edge research. The ideal candidate will work towards publishing his findings in a competitive conference in machine learning.

HOW TO PROCEED

Please contact me on my personal email asap. Make sure to attach your c.v., as well as any other information you think will help me determine your experience and expertise (e.g., a recommendation letter, copies of your grades, links to previous projects, blog, link to github account).