Master’s project: 'Mind Reading' with MEG and Machine Learning

Abstract

Can we read the thoughts by looking at the neural signals produced by the brain?

Reading complex thoughts is clearly still beyond the scope of our technology (for better or for worse). In this Master’s project our aim is much simpler: we wish to predict whether the person is thinking of a noun (or object-like imagery) as opposed to verb (or action-like imagery).

To this end, we will use magnetoencephalography (MEG), a technology designed to read magnetic fields created by the current produced by neurons in real-time. The signal obtained will then be analysed with a Machine Learning algorithm to classify noun vs. verb condition.

The results show that we can indeed 'read minds'. In particular, we observe specific patterns of activation in what is known as Episodic Memory Network which is thought to process episodic memories.

Top level workflow

• Preprocessing (per subject)
• Dataset generation (per cortical area + sensor space)
• Adding conditions (per condition type)
• MVPA analyses (per analysis)

graph LR;
  preprocess --> dataset;
  dataset --> conditions;
  conditions --> analyses;

Loading

Per subject (preprocessing.py)

• create_directories()
  • => bunch of directories
• read_raw()
  • => raw object
• downsample() (optional)
  • => downsampled raw object
• filter() (optional)
  • => filtered raw object
• remove_artifacts() (optional)
  • => raw object without artifacts
• epoch()
  • => epoch file
  • => sensor space data
• source_localize()
  • => source space data

graph LR;
    create["create_directories(...)"] --> read_raw;
    read_raw["read_raw(...)"] --> downsample;
    downsample(["downsample(...)"]) --> filter;
    filter(["filter(...)"]) --> remove_artifacts;
    remove_artifacts(["remove_artifacts(...)"]) --> epoch;
    epoch["epoch(...)"] --> source_localize;
    source_localize["source_localize(...)"];
    
    classDef parallel fill: #990000, stroke: #333, stroke-width: 3px, font-size:10px;
    

Loading

Per cortical area (dataset.py)

• generate_area_data()/generate_area_data_mmap()

NB: round nodes are optional