Kirill Mavreshko

Software Developer (as a freelancer)

- Design and development of both of backend and partially frontend (for admin site) parts of the portal. Did some optimizations to withstand high traffic. Technologies: Python, Django, JavaScript, MySQL, PostgreSQL, Nginx, HTML, CSS, SVN, Trac.

Gzt.ru has been closed in 2011.

Entrepreneur

Co-founder.
- Was responsible for the development of both the backend and the frontend parts of the portal. Integrated portal's database with multiple real estate agencies. Technologies: Python, Django, PostgreSQL.
- Website maintenance & support

Lead Software Developer

Backend Web developer. Technologies: Python, Django, PostgreSQL, Linux, Mercurial.

Software Developer

- Design and implementation of several WEB-oriented systems using Perl, PHP, .NET, Java and MySQL with PostgreSQL as DB backend. OS: Linux (mostly) / Windows

Keras-transformer

Keras-transformer is a library implementing nuts and bolts for building (Universal) Transformer models using Keras. It allows you to assemble a multi-step Transformer model in a flexible way.

The library supports positional encoding and embeddings, attention masking, memory-compressed attention, ACT (adaptive computation time). All pieces of the model (like self-attention, activation function, layer normalization) are available as Keras layers, so, if necessary, you can build your version of Transformer, by re-arranging them differently or replacing some of them.

The (Universal) Transformer is a deep learning architecture described in arguably one of the most impressive DL papers of 2017 and 2018: the "Attention is All you need" and the "Universal Transformers" by Google Brain team.

Their authors brought the idea of recurrent multi-head self-attention, which has inspired a big wave of new research models that keep coming ever since, demonstrating new state-of-the-art results in many Natural Language Processing tasks, including translation, parsing, question answering, and even algorithmic tasks.

My impression is that indeed, unlike already classical recurrent neural networks, Transformer trains much faster measured both as the time per epoch and the wall clock time. It's also capable of efficiently handling multiple long-term dependencies in texts.

For instance, when applied to text generation, Transformer creates more coherent stories, which don't degrade in quality with the growth of their length, as it is typically the case with the recurrent networks.

Kenlp

The implementation of paper "Mimicking Word Embeddings using Subword RNNs". MIMICK allows to avoid OOV (out of vocabulary) problem by imitating the original pre-trained word embeddings using small character-based model, thus making < UNK > word embeddings unnecessary.

Benefits:

• very low resource requirements (in comparison with FastText, for example)
• significant boost in accuracy due to the absence of < UNK > in your training data

The demonstration of pre-trained model which can mimic word embeddings for OOV words and show which of the known words are closest to the produced vector:

The word "trintiful" is fictional but MIMICK "understands" that this is probably some kind of adjective. It assumed that "unfrogable" is a negating adjective. It was also able to guess that "Kaa" looks like an asian name, and "Karll" is something european. "abroktose" turns out to be something vaguely scientific, probably about chemistry or biology, which also makes sense.

Kerl

KERL is a collection of various Reinforcement Learning algorithms and related techniques implemented purely using Keras.

The goal of the project is to create implementations of state-of-the-art RL algorithms as well as a platform for developing and testing new ones, yet keep the code simple and portable thanks to Keras and its ability to use various backends. This makes KERL very similar to OpenAI Baselines, only with focus on Keras.

What works in KERL:

• Advantage Actor-Critic (A2C) [original paper]
• Proximal Policy Optimization (PPO) [original paper]
• World Models (WM) [original paper]

All algorithms support adaptive normalization of returns Pop-Art, described in DeepMind's paper "Learning values across many orders of magnitude". This greatly simplifies the training, often making it possible to just throw the algorithm at a task and get a decent result.

With KERL you can quickly train various agents to play Atari games from pixels and dive into details of their implementation. Here's an example of such agent trained with KERL (youtube video): Deep RL A2C Agent Playing Ms. Pacman

Limitations:
Currently KERL does not support continuous control tasks and so far was tested only on various Atari games supported by The Arcade Learning Environment via OpenAI Gym.

CartGP

CartGP is a very simple and minimalistic C++/Python library implementing Cartesian Genetic Programming (CGP). The library currently supports classic form of CGP where nodes are arranged into a grid and no recurrent connections are allowed.

Check this jupyter notebook to see how to use the library from Python.

Mittleider Weekly-Feed Mix Calculator

I love gardening. After months of studying how to grow plants the right way, I have created a calculator which uses mathematical optimization (linear programming) to make sure my mix of fertilizers always fits recommendations from experts in plant nutrition and contains everything necessary for plant's healthy growth.

With this calculator, you can quickly create a fully balanced Dr. Jacob Mittleider Weekly-Feed mix using virtually any sorts of fertilizers locally available to you.

The extra benefit of this tool is that it can guide the gardener step-by-step through the process of creating the mix by pointing on problems with the existing recipe.