Tolga Recep Uçar |
|
|
Mathematics @ Koc University
tucar24@ku.edu.tr |
|
TUBITAK granted research (2247-C.) Supervised by Prof. Dr. Mehmet Fatih Amasyalı. We propose an alternative VAE architecture inspired by the inner workings of the ubiquitous Transformer. We also diagnose the condition of posterior collapse on the latent space and aggressively train the encoder to create an efficient model overcoming this. Finally the architecture is trained on a large Turkish corpora and tested on various tasks such as sentiment analysis and text classification. Publicly available here.
(Accepted.) We propose a numerical algorithm based on neural networks for solving linear and nonlinear ordinary differential equations. We combine the representational power of orthogonal space of Jacobi polynomials and multilayer perceptrons. Our method presents low-cost and highly accurate approximations with basic training. Some of the tinkering before the actual architecture was designed can be found here.
(Accepted for on-site oral presentation and print.)
Abstract:
We propose machine learning as a
new method for solving Abel type differential
equations which is an important class of
differential equations modeling magnetostatic
problems and fluid dynamics. Our method does
not rely on extensive formal numeric
computing. It is a data-driven algorithm, hence
a significant approach as data is more available
each day. Our approximation function is a
multilayer perceptron and we adjust parameters
by backpropagation algorithm. We compare our
results with some well-known approximate
results.
Graduate Research and Teaching Assistant: Research on numerical linear algebra and eigenvalue solvers.
Assisting and grading Math 102 - Calculus course for undergraduates.
Data Engineer: Cleaning data and structuring it for analysis and
modeling at Turkish Airlines. Developing supervised and unsupervised learning solutions for
component planning problems.
(and I was a Freelance Python Tutor before this.)