-
Notifications
You must be signed in to change notification settings - Fork 14
Glossary
Minerva is an educational project that lets Users learn advanced machine learning via reproducing and understanding winning solutions to Kaggle challenges.
Problem is an entire end-to-end pipeline (usually quite complex) that constitutes solution to the Kaggle challenge. For example, whales problem implements Right Whales Recognition winning solution.
Learning objectives show what practical machine learning techniques and tricks will be used. User experiments with various parts of the solution pipeline in order to learn how to tune the model, preprocess the data, validate the model or even exploit data leaks. All of these activities leads to the re-implementation of the top solution.
User starts working on a problem by opening the welcome notebook (for example, Right Whales Recognition). Each problem is implemented in a clear way and divided into logical pieces. These pieces are called tasks - User's main point of interest. User is encouraged to solve all the tasks. User can work on tasks in any order and eventually re-implement a complex data scientific pipeline.
Task contains a single processing step, such as a ConvNet implementation or a post-processing function. Each task is defined in the corresponding Jupyter Notebook (for example, task 1 in the Right Whales Recognition problem), where User can find the description of the task, learning objectives, what is expected as a solution and some auxiliary information that should help in the implementation process.
Technically, User works on the task by implementing the solution function, stub of which is already in the notebook. When the solution is ready, User submits the task and the corresponding part of the pipeline is replaced by the User's implementation. Recall that the entire pipeline is pre-implemented by Minerva team.
User submits the task by running a single command. The structure of this command is explained in detail in the User Guide.
Currently, feedback is provided in a binary way: if the final score is very close to the default pipeline implementation, User receives an appropriate praise, if it is not, User is asked to try again.
Neptune is a machine learning lab which gives you an access to powerful external resources and allows you to easily track the progress of your experiments. Minerva is fully integrated with Neptune, however registering in Neptune is not required at any point.