Martin Mirakyan

A graph neural network approach to approximate betweenness-centrality ranking in Graphs. Experiments on real-world and synthetic datasets show that our algorithm is faster in inference and requires several times fewer resources and time to train.

This work is submitted as a journal paper.

EPE-NAS with DARTS and for faster neural architecture search.

PyTorch Lightning implementation of differentiable neural architecture search with benchmarks on both DARTS search space and NasBench-201 search space. The approach is tested on several datasets, including CIFAR-10, CIFAR-100, MNIST, Fashion-MNIST, and Imagenet.

This library provides a low-level API to perform operations directly on neurons to utilize individual neurons in a neural network better. The current neural network library implementations are concentrated on layer-based operations.

I published a report on the reproducibility of a DIIN neural network architecture for natural language inference using Keras. The code is open source and available on GitHub.

• https://github.com/YerevaNN/DIIN-in-Keras

I published a blog post on the implementation details and reproducibility of the R-net paper.

• http://yerevann.github.io/2017/08/25/challenges-of-reproducing-r-net-neural-network-using-keras/
• https://github.com/YerevaNN/R-Net-in-keras

I developed an end-to-end pipeline for extracting "linguistically accurate" morphological vectors for given words or sentences. The system is based on FastText and COMBO. Experiments are tested and evaluated in the Russian language.

Source Code:
• https://github.com/MartinXPN/word2morph
• https://github.com/MartinXPN/morph2vec
• https://github.com/MartinXPN/sentence2tags
• https://github.com/MartinXPN/word2morph2vec
• https://github.com/MartinXPN/ag-annotation

I architected a convolutional neural network with custom attention interaction and used a bidirectional recurrent neural network as a baseline.

Source Code
• https://github.com/YerevaNN/RelationClassification

Image Coloring uses neural networks, and this implementation is a Wasserstein GAN for turning greyscale images into colorful ones.

Three Experiments:
• A simple approach: colorize a greyscale image in the lab space. Given the L channel as an input, predict AB channels.
• Classification: the AB space was divided into chunks of colors and treated as a classification problem.
• Wasserstein GAN: used generative adversarial networks to predict AB channels with both first and second approaches.

Spell Checker is an app designed for editing and correcting texts. The app works offline and supports several languages, including Armenian, English, and Russian. The app currently has more than 100,000 downloads.