problem statement is to find representations that are invariant to inter and intra subject diffrences
- reduce noise associated with EEG data is key → find resources
task at hand is mental load classification task
Procedure is as follows
- EEG activity data trasnformed into topology-preserving multi-spectral images: captures image from specific wavelength
- Train on recurrent convolutional network based on video classification techniques to learn about the images themselves
- preserves spatial, spectral, and temporal structure of EEG → leads to finding features that are less sensistive to variations and distortions within each dimension
- RNNs found to have better and improved classification results
Dataset is key issue (found on Kaggle lmao)
deep belief network and ConvNets have been used to learn representations from MRI and same approach can be used for neuroimaging
Convolution and recurrent neural networks used to extract representations from EEG time series → studies demonstrated potential benefits of downscaling data and applying this approach → goal was to preserve structure of EEG data within space, time, and frequency
EEG workings
- uses changes in electrical voltage acros the scalp induced by cortical activity:
- Restricted Boltzman Machine layers to a deep belief network and using supervised pretraining results in networks that can learn increasingly complex representations of the data and achieve considerable accuracy increase as compared to other classifiers
- problem was framed similarly to that of multi-channel speech obtained from several different microphones or in this case, the electrodes to record signals from speakers or cortical regions
- goal is to learn the discriminative manifold between different states and corresponding EEG readings
- procedure involved turning EEG features into multi-dimensional tensor which retains structure of data throughout learning process over the traditional vector usecase
- train on deep recurrent-convolutional neural network architectures inspired by video classification to learn time series data
- use ConvNets to extract spatial and sepctral invariant rep. from each frame data
- adopt LSTM network to extract temporal patterns in frame sequence
- this approach reduced classification error to 8.9% from 15.3%