Abstracts Track 2021

Area 1 - Complexity in Biology and Biomedical Engineering

Nr: 3

Modeling of Multicellular Colonies Functioning under Programmed Cell Death (apoptosis)


Alina Shipitsyna, Timofey Lomonosov and Fuad Aleskerov

Abstract: Apoptosis is a form of programmed cell death. Cells dying from apoptosis are known to be absorbed by neighboring cells in some cases. In this paper, apoptosis is modeled using a differential equations and an assumption that part of dead cells are absorbed by their neighbors. Also, modeling was conducted with taking into account both external and internal pathways of apoptosis. Using numerical methods, a study of the behavior of the colony is carried out for various parameters such as initial level of nutrients, initial number of healthy cells, coefficient of absorption of dead cells by healthy cells in colony. With a sufficient amount of nutrients, a rapid growth of the colony and then the establishment of some almost constant number of cells were observed. Number of healthy cells under the lack of the nutrients was found to decrease at the first time and then begin to increase as the colony adapts.

Area 2 - Complexity in AI/Edge/Fog/High-Performance Computing

Nr: 4

Generalizing Neural Autoregressive Quantum States


David R. Vivas Ordóñez, John Henry Reina Estupiñán and Javier Madroñero Pabón

Abstract: Generative neural-network models such as Restricted Boltzmann Machines (RBMs) have recently faced successful applications in quantum physics as powerful approximators of many-body wavefunctions, regarded in the literature as Neural Quantum States. Despite this, these models are limited in by the intractable nature of their probability density, and have traditionally required the use of Markov-Chain Montecarlo methods for sampling. Most recently, a novel approach based on the use of deep fed-forward autoregressive convolutional neural-networks with direct sampling capabilities has been used to efficiently approximate the ground-state wavefunction of large quantum many-body systems. Here, we illustrate how such an approach generalizes to a wider spectrum of NQS architectures.

Area 3 - Complexity in Informatics and Networking

Nr: 5

Higher-order Spectral Clustering


Andrei Bobu, Konstantin Avrachenkov and Maximilien Dreveton

Abstract: Graph clustering is the task of identifying groups of tightly connected nodes in a graph. This is a widely studied unsupervised learning problem having a wide range of applications in computer science, statistics, biology, economy or social sciences. One of the most popular methods of graph clustering is spectral clustering (von Luxburg, 2007). If the graph has to be splitted into two communities, the simplest version of such an algorithm uses for partitioning the eigenvector associated with the second smallest eigenvalue of the graph's Laplacian matrix (the so-called Fiedler vector (Fiedler, 1975)). However, although this method is rather efficient both theoretically and practically for many graphs (for instance, in the Stochastic Block Model (Abbe, 2017)), it can be heavily handicapped by geometric structure in a network. Such an example is the Geometric Block Model (GBM) recently introduced in (Galhotra et al., 2018). We present an effective generalization of standard spectral clustering, which we call higher-order spectral clustering. It is close to the classical spectral clustering method but uses for partitioning the eigenvector of the adjacency matrix associated with an eigenvalue of higher order. This generalization provides weak consistency for a new model of geometric graphs, called Soft Geometric Block Model (SGBM). This result is established in three steps. Firstly, we charaterize the limiting spectrum of the adjacency matrix extending the result of (Bordenave, 2009). Secondly, we take an eigenvalue closest to some 'optimal' value and show that this eigenvalue is separated in the limiting spectrum under certain conditions. Finally, using Kahan-Parlett-Jiang theorem, we prove that the eigenvector corresponding to the considered eigenvalue cannot be far from the vector recovering the true community structure by the signs of its coordinates. A small adjustment of the algorithm provides strong consistency. This adjustment is called local improvement step, which consists of counting neighbours of each node in the supposed communities. We also show that higher-order spectral clustering is effective in numerical experiments even for graphs of modest size. In particular, in the case of dense GBM graphs, it outperforms numerically the algorithm of (Galhotra et al., 2018). For more details and proofs, we refer to our preprint (Avrachenkov et al., 2020). Ulrike von Luxburg. A tutorial on spectral clustering. Statistics and Computing, 17(4):395–416, 2007. Miroslav Fiedler. A property of eigenvectors of nonnegative symmetric matrices and its application to graph theory. Czechoslovak Mathematical Journal, 25(4):619–633, 1975. Emmanuel Abbe. Community detection and stochastic block models: recent developments. The Journal of Machine Learning Research, 18(1):6446–6531, 2017. Sainyam Galhotra, Arya Mazumdar, Soumyabrata Pal, and Barna Saha. The geometric block model. In Thirty-Second AAAI Conference on Artificial Intelligence, 2018. Charles Bordenave. Eigenvalues of Euclidean random matrices. Random Structures & Algorithms, 33(4):515–532, 2008. Konstantin Avrachenkov, Andrei Bobu and Maximilien Dreveton. https://arxiv.org/pdf/2009.11353.pdf

Area 4 - Complexity in Risk and Predictive Modeling

Nr: 7

Sunburst: Protecting the Full Software Delivery Pipeline


Tony Hadfield

Abstract: We are going to take a look at the full software deliver pipeline. What can you do to prevent the next Sunburst? Looking at Sunburst, the ultimate delivery of malware was introduced into the final software package due to malware running on a build machine that was modifying source at build time. Protecting against an attack like this can be a bit of a whack-a-mole situation. If we just focus on the malware, the attacker would be on the hunt for other ways to exploit the system. Instead of focusing just on how to prevent malware in your software delivery pipeline, let's take a look at the controls that should be in place to ensure you are delivering software your consumers can trust. We'll some important areas that should be audited and controlled, including the following: - Source code commits, are you requiring your developers to sign their commits? - Does your build pipeline include source code analysis? - How do you validate and lock down a build system? - For artifact signing - how do you control who and where signing happens with your keys? - For artifact signing, how do you know and approve what is signed? - If something goes wrong, how do you revoke a key, and what will be impacted? For more info on this topic, see the blog post here: https://venafi.com/blog/sunburst-code-signing-was-problem-it-should-have-been-solution