I am a research scientist (Ambizione fellow) at the LTS2 lab in EPFL, Switzerland.

My research focuses on the foundations and applications of machine learning to structured problems. I aim to find ways to exploit (relational, group, constraint) information, with the ultimate goal of designing algorithms that can learn from fewer data. I am also fascinated by the theoretical analysis of neural networks and in using them to solve bio-engineering problems (especially protein design).

Selected recent works

Y. Dong, J.B. Cordonnier, A. Loukas. Attention is Not All You Need: Pure Attention Loses Rank Doubly Exponentially with Depth. Oral at ICML 2021. (preprint, bibtex, blogpost)

N. Karalias, A. Loukas. Erdos Goes Neural: an Unsupervised Learning Framework for Combinatorial Optimization on Graphs. Oral at NeurIPS 2020. (paper, bibtex, video starts at 36:54, blogpost, slides, talk at Yale)

A. Loukas. How hard is to distinguish graphs with graph neural networks. NeurIPS 2020. (paper, bibtex, video, blogpost)

C. Vignac, A. Loukas, P. Frossard. Building powerful and equivariant graph neural networks with structural message-passing. NeurIPS 2020 (paper, bibtex, video, blogpost)

A. Loukas. What graph neural networks cannot learn: depth vs width. ICLR 2020. (paper, bibtex, blogpost, 5min-presentation)

JB Cordonnier, A Loukas, M. Jaggi. On the relationship between self-attention and convolution. ICLR 2020. (paper, bibtex, blogpost, code, interactive website)

JB Cordonnier, A Loukas. Extrapolating paths with graph neural networks. IJCAI 2019. (preprint, bibtex, blogpost, code)

A Loukas. Graph reduction with spectral and cut guarantees. JMLR 2019. (paperbibtexcodeblogpost)

A Loukas. How close are the eigenvectors and eigenvalues of the sample and actual covariance matrices? ICML 2017. (paperbibtexblogpost)

Additional information

List of publications

Contact details

Social media

Please consult google scholar.

Drop me an email at “firstname.lastname@epfl.ch”.

Catch me on twitter, researchgate, or linkedin.

Andreas