2018/06/20 by Joos Behncke, Robin Tibor Schirrmeister, Behncke, Joos +13
Computer Science · Neuroscience · #Blind Source Separation Techniques #EEG and Brain-Computer Interfaces #FOS: Biological sciences #FOS: Computer and information sciences #Human-Computer Interaction (cs.HC) #Neural dynamics and brain function #Neurons and Cognition (q-bio.NC)
paper · pdf · doi:10.48550/arxiv.1806.09532
openalex publication_date 2018/06/20 · openalex created_date 2022/08/18 · openalex updated_date 2026/07/28
When it comes to the classification of brain signals in real-life\napplications, the training and the prediction data are often described by\ndifferent distributions. Furthermore, diverse data sets, e.g., recorded from\nvarious subjects or tasks, can even exhibit distinct feature spaces. The fact\nthat data that have to be classified are often only available in small amounts\nreinforces the need for techniques to generalize learned information, as\nperformances of brain-computer interfaces (BCIs) are enhanced by increasing\nquantity of available data. In this paper, we apply transfer learning to a\nframework based on deep convolutional neural networks (deep ConvNets) to prove\nthe transferability of learned patterns in error-related brain signals across\ndifferent tasks. The experiments described in this paper demonstrate the\nusefulness of transfer learning, especially improving performances when only\nlittle data can be used to distinguish between erroneous and correct\nrealization of a task. This effect could be delimited from a transfer of merely\ngeneral brain signal characteristics, underlining the transfer of\nerror-specific information. Furthermore, we could extract similar patterns in\ntime-frequency analyses in identical channels, leading to selective high signal\ncorrelations between the two different paradigms. Classification on the\nintracranial data yields in median accuracies up to (81.50 \± 9.49) , %.\nDecoding on only 10 % of the data without pre-training reaches performances\nof (54.76 \± 3.56) , %, compared to (64.95 \± 0.79) , % with\npre-training.\n