Deep Learning & Convolutional Neural Network Visualization

 

Visualization of trained convolutional neural network

Visualization of trained convolutional neural networks, showing the correlation between changes in band power of input signals and changes in predictions of convolutional neural networks (see Schirrmeister et. al (2017) for more details)

 

Deep learning has revolutionized fields such as computer vision and speech recognition, yielding unprecedented performance in many machine learning tasks. Deep learning is also increasingly applied for medical tasks, prominent examples are convolutional neural networks (ConvNets) that have diagnosed skin cancer with higher accuracies than actual dermatologists (Esteva et al., 2016) or that have been applied to medical images from eye examinations (De Fauw et al., 2017). We apply deep learning to the task of brain-signal decoding:  Concretely, we use convolutional neural networks on EEG signals. Our focus is on adapting the network architectures and training strategies to the particularities of EEG decoding tasks and creating visualizations to make the trained models interpretable.

In our work, we have already shown convolutional neural networks can reach competitive decoding accuracies for decoding movements from human EEG. In the same work, we also developed visualization methods that show how band-power features affect the trained ConvNets predictions, yielding highly plausible spatial maps. In the future, in collaboration with machine learning researchers at the University of Freiburg, we want to improve the process of doing research with ConvNets by using automatic hyperparameter optimization, apply ConvNets to larger and more diverse datasets and create more methods that help neuroscientists and others get an understanding of what features the ConvNets are extracting from the brain signals.

 

  • De Fauw, J., Keane, P., Tomasev, N., Visentin, D., van den Driessche, G., Johnson, M., … Cornebise, J. (2016). Automated analysis of retinal imaging using machine learning techniques for computer vision. F1000Research, 5, 1573. http://doi.org/10.12688/f1000research.8996.1
  • Esteva, A., Kuprel, B., Novoa, R. A., Ko, J., Swetter, S. M., Blau, H. M., & Thrun, S. (2017). Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639), 115-118.
  • Schirrmeister, R. T., Springenberg, J. T., Fiederer, L. D. J., Glasstetter, M., Eggensperger, K., Tangermann, M., ... & Ball, T. (2017). Deep learning with convolutional neural networks for brain mapping and decoding of movement-related information from the human EEG. arXiv preprint arXiv:1703.05051.
Personal tools