ML: Predict sequence of values

Let's say we have nice a line built up of two damped oscillators as it displayed on the picture.

What if I say it's possible to predict 700 values of this line using just:

  • 30 data points for feeding neural network(which is just half-period)
  • using just two fully-connected layers (hence it's not deep network)
  • having just three neurons in whole network

Read more

First experiments with ML using Keras

I'm still exploring the new world of machine learning.

Recently I discovered few interesting things about it:

  • It's not only for math nerds. A lot of tools are on market, you just have to understand how to use them
  • There are a lot of application of ML, it might be hard to see that from the first sight but it because it's a completely new way of solving problems you probably never thought about
  • Investors LOVE it, any project which has AI/ML in the description make it more attractive for investments
  • It's the next Big Thing as you might see already
  • A lot of human's mechanic work can be replaced with AI
  • In terms of scientific researches I see two kind of them: 1) Very deep math used for building ML tools 2) Adaptation of existing tools, model configurations, best practices for new applications

Hence, I started experimenting with it. First of them was Prisma-like chat bot for Facebook which uses open source implementation of neural style transfer algorithm. Then I made AI-powered visa sponsored job search engine which discovers companies' websites and walk over job listings to extract vacancies. It uses CNN for binary classification to recognise job listings.

Read more

Игры с neural style transfer

Последний месяц для саморазвлечения в качестве сайд-проекта играюсь с neural style transfer - нейронные сети, переносящие стиль с картинки на картинку. Многие знают это как приложение Prisma. Видимо создатели тоже читают блог creativeai :)

Первая реализация появилась больше 8 месяцев назад и доступна на github jcjohnson/neural-style - она написанна на lua, всё замечательно, кроме того что работает она вечность (и жрёт почти бесконечность памяти). На Digital Ocean с 16 swap генерация картинки 450x200 заняла около 2 минут, сейчас уже не помню.

Read more