Elod Pal Csirmaz
(he/him)
I work as a lead software architect and have a keen interest in artificial intelligence, algebra, filesystems and computer science in general, as well as in creative writing and literary criticism. I live in the UK with my husband and two children.
Computing
-
Attempting the impossible: Enumerating extremal submodular functions for n=6
By Elod Csirmaz and Laszlo Csirmaz.
Submodular functions have applications in such diverse areas as information inequalities, operational research, combinatorial optimization and social sciences, and they have also found fundamental applications in game theory and machine learning. Enumerating the extremal submodular functions defined on subsets of a fixed base set has only been done for base sets up to five elements. This paper reports the results of attempting to generate all such functions on a six-element base set. Using improved tools from polyhedral geometry, we have computed 360 billion of them, and provide the first reasonable estimate of their total number, which is expected to be between 1,000 and 10,000 times this number. We also present insights into the symmetries of these functions, as well as two new inequality insertion orders that allow processing them further than before.
Available on arXiv ». The related code is available on GitHub ». The partial result set can be found on Zenodo ».
-
Synchronizing Many Filesystems in Near Linear Time
By Elod Csirmaz and Laszlo Csirmaz, in Future Internet 2023; 15(6):198.
Finding a provably correct subquadratic synchronization algorithm for many filesystem replicas is one of the main theoretical problems in Operational Transformation (OT) and Conflict-free Replicated Data Types (CRDT) frameworks. Based on the Algebraic Theory of Filesystems, which incorporates non-commutative filesystem commands natively, we developed and built a proof-of-concept implementation of an algorithm which synchronizes an arbitrary number of replicas. It is provably correct, and has linear complexity after an initial sorting phase. It works by identifying conflicting command pairs and requesting one of them to be removed; can be guided to reach any of the theoretically possible synchronized states.
Also available on arXiv ». See sample implementation of the filesystem synchronization algorithms on GitHub »
-
Data Synchronization: A Complete Theoretical Solution for Filesystems
By Elod Csirmaz and Laszlo Csirmaz, in Future Internet 2022; 14(11):344.
Data reconciliation in general, and filesystem synchronization in particular, lacks rigorous theoretical foundation. This paper presents, for the first time, a complete analysis of synchronization for two replicas of a theoretical filesystem. While existing synchronizers are operation-based as they define, using some rationale or heuristics, how conflicts are to be resolved without considering the effect of the resolution on subsequent conflicts, our approach is declaration- based: we define what constitutes the resolution of all conflicts, and we prove the existence of sequences of commands which convert the replicas into a common synchronized state.
All possible synchronized states are described by a partial order on the conflicts describing the order in which they need to be resolved, as well as the effect of each decision on subsequent conflicts. Using this classification, the outcomes of different conflict resolution policies can be investigated easily. Also available on arXiv »
-
Interpreting Neural Networks by Reducing Nonlinearities during Training A short paper and sample code demonstrating a simple solution that makes it possible to extract rules from a neural network that employs Parametric Rectified Linear Units (PReLUs). We introduce a force, applied in parallel to backpropagation, that aims to reduce PReLUs into the identity function, which then causes the neural network to collapse into a smaller system of linear functions and inequalities suitable for review or use by human decision makers.
Read the paper »
More on GitHub »