MLIS - machine learning in seconds

• 

  Login page, click login with facebook to login

• 

  Main page with list of problems

• 

  Every problem have problem description and code editor. You can see problem ranking, submit, see run results and load code from file

• 

  For submission, you can see results of the run, if it pass tests and if not passed why it failed

• 

  For problem, you can see rating based on problem metric

• 

  You can also see global rating based on performance on all problems

• 

  If you need mentorship, you can apply for a class and get mentor and classmates

• 

  Once you in a class, you can talk with classmates and mentor in chat

• 

  If you want to study alone, you can start with easy tutorial task, that help you get started via Google Colab Notepad

• 

  Mobile friendly interface, try https://www.mlisjudge.com on the phone and start learning machine learning on your phone

Inspiration

One day, I decided to study machine learning. Soon I realized that I am not studying most of the time but waiting for my neural network model to learn. Waiting for minutes, hours and sometimes days was boring and not productive. I wanted to see results in seconds and iterate quickly. I wanted to study effectively. So I created machine learning in seconds (MLIS) platform. Here you consolidate your theoretical knowledge with practical tasks that can be done in under 2 seconds. You can experiment and study more effectively. We simplify real problems in such a way so you still can work with state of the art techniques for a neural network and you don't need to wait for hours on every iteration. You can solve tasks on your computer or on your mobile phone. MLIS - machine learning in seconds.

What it does

MLIS is machine learning in seconds. It is a platform where you can compete optimizing neural networks and learning algorithms for different problems, with the only constraint: a neural network should be trained in under 2 seconds. If you are new to neural networks, you can study them on our platform, where you would be given a weekly task, a mentor and a community of classmates.

How I built it

There are 4 main parts:

• Client - react/relay web/mobile application
• Server - GraphQL server implemented on NodeJS
• Agent - NodeJS agent responsible for running dockerized training of user's PyTorch model written in Python
• Problem set - I usually read some ML paper with some cool technic, then I create a problem that can be solved using such technic in under to 2 seconds with Pytorch, then I let people solve it if there is some hack, I try to update problem

Challenges I ran into

• Installation is a real pain for teaching machine learning, when people are not sure whether they want to study it or not, installation killing all the fun
• GPU is expensive

Accomplishments that I'm proud of

When I started a lot of people was telling me that a neural network can't learn in under 2 seconds. I'm proud that I was able to come up with problems/datasets which can be learned in under 2 seconds and you still need state of the art technics to solve them (batch norm, structure/convolution, attention, etc) and now you can do it from your mobile home during your commute to work.

What I learned

While mentoring people on MLIS, I learned that often students would have the right solution, except one line or even one character. It is extremely hard to debug neural networks. Eventually, I decided to collect such solution with small errors for competitions.

What's next for MLIS - machine learning in seconds

• Move mentor/student communication from Messenger Group to the web app
• More classes: Advanced: Speach recognition, Advanced: Graph neural networks, etc
• Find sponsor for GPU support
• Weekly competitions where you need to fix a solution changing only n character/lines.
• Automate mentorship for Basic class
• Code free machine learning challenges for children

Built With

• amazon-web-services
• bootstrap
• docker
• graphql
• node.js
• postgresql
• python
• pytorch
• react
• relay