We’re decoding brain signals using Azure Machine Learning

For all those fans of Data Science, Microsoft have recently released Cortana Intelligence Competitions. A gamification feature of the Cortana Intelligence Suite.

The first competition has started and it’s a beauty! It involves decoding brain signals in an attempt to predict whether a person has been shown one image or another based on the nature of the activity in their brain. The signals are picked up via 64 electrodes fitted to the patients’ head during the testing phase.

Incredible, no…

Brain data

When starting, you are supplied with data collected from 4 patients who, whilst connected to 64 electrodes around the head, have each been shown 300 grey-scale images of either a house or face.

In this brain training data, you can decipher which of the 2 stimulus types the patient has been shown, along with the brain signals, and this brain data is what you then use to build predictive models for predictions on unseen data based on brain signal activity.

It’s the level of predicting we will be able to do, that is so exciting.

The Data Modelling Process

This competition starts you off with an example experiment into which you add your own modules for pre-processing (to clean the electrical signals, for example), data splitting (for training and cross validation), model training and scoring.

Contestants need to follow a typical data science iterative process of continually refining the data pre-processing, model selection and parameter tuning:

Testing the Machine Learning

Once happy with your model you can submit it for testing (a maximum of 3 times per day) against a dataset which you cannot see, and you get an accuracy score on a publicly available leader board showing the percentage your model correctly predicted.

One data set is held back for when the competition completes on the morning of the 1st July 2016 and each contestant’s 5 most accurate models are scored against this dataset to decide the final positions.

‘Team Pixie’

Three members of the RedPixie data science team are competing independently in the competition and are holding positions 2nd, 9th and 12th out of 276 contestants at the time of writing.

There a still a couple of months to go and these competitions tend to hot up closer to the final date but we are going to try hang on in there at the top of the table if we can, free time and day jobs permitting.

We’ll keep you updated with our progress.

Written by Stuart Anderson | Principal Consultant, RedPixie | See his LinkedIn Profile | Follow him on Twitter