Learning to recall dreams from dreamers

This is an attempt to classify people who had a high dream recall rate from those who didn't based on their EEG data recorded during sleep using a deep convolutional network architecture.
More about me

Aims

• Train a deep convolutional architecture to discriminate people with high dream recall rate (dreamers) from those with low dream recall rate (non-dreamers).
• Observe the decoding (classification) accuracies for different sleep stages to identify key sleep stage containing differences
• Visualize the features extracted by the CNN to classify the two groups
• Try to train on identifying individual subjects - maybe used as person identification tool

Dataset

Courtesy - CoCo Lab

• 36 subjects - 18 dreamers and 18 non-dreamers
• Sleep trials of 30 sec interval sampled at 1000 Hz
• 19 EEG electrodes
• Sleep trials segregated based on sleep stage (S1, S2, REM and SWS) as annotated by an expert

Tools/Language used

• MATLAB
• Torch (Lua)

Data Processing and Structuring

• Data downsampled to 200 Hz
• Check for anti-aliasing (Thanks Andrew)
• Sleep trials split into segments of 5 sec long each
• Data dimension structured to be 1 X 19 X 1000 to pass to the CNN

Deep Learning architecture

• Train network on SWS sleep data. 80% training data from bag of sleep segments, 20% for validaton
• Train network on 34 subjects and test on remaining 2 subjects -> accuracy not good enough for 14 folds
• Use REM and S2 data. S1 data may not be enough!! :/
• Train a subject identifier
• Choose parameters for good subject classifier
• Confusion matrix for subject prediction
• Try subj prediction with lesser training data -> Better accuracy with more filters (more params)
• Too many params in network, yet better accuracy -> Increasing linear layers improves subject ID, but affects dreamers-nonDreamers, hence more linear layers allow subject specific features to be learnt.

Visualizing features

• Use GAP
• Use gradCAM/deep dream
• Use ccCAM (method being developed)

Current issues

• Good accuracy at sleep segment level, but not at subject level
• Write code for identifying subjects

Deliverables for Brainhack school

• Develop a re-usable framework in Torch -> Training and Data preparation files added with some comments, descriptions added in Readme file in Codes folder
• Port the code to PyTorch and make a notebook
• Write down a blog post detailing challenges faced and tips/tricks to solve them

Further scope

• Interpret neurological basis of extracted features
• Learning how people dream or what affects dream recall
• Extend to bigger datasets of M/EEG recordings

Important Resources

• Visualizing CNN decision in automated sleep scoring - Link