Framework for Training BCI System Ilya Kuzovkin MTAT.03.277 Research Seminar in Data Mining University of Tartu, 2012
Outline � Background � The Task � Usual approach & difficulties � Our attempt to improve things � Results
Background Neurons transmit information via electrical signaling.
Background Different areas of the scalp have different electric potentials We put electrodes and register these differences
The Task Distinguish different mental states and map them to actions.
Usual approach & difficulties Signal Processing
Usual approach & difficulties Machine Learning Instance (row) joint spectrums from all channels at the moment of time Attribute (column) certain frequency power on certain channel Class specifies the instance label
Usual approach & difficulties � Signal is noisy Skull � Skin � Hair � We can’t measure signal at it’s origins � � Lack of data? No. � Bad algorithms? No. � Signal is not consistent for same mental state � Test subject does not know how different his thoughts are in terms of resulting electrical signal
Our attempt to improve things General Idea � We want to find a way to make learning process duplex: not only machine will learn from user input, but user will learn and adapt his behavior according to feedback he receives from a machine � This can be formulated as establishing a communication protocol between two agents during the interaction � Goal is to find set of messages which can be clearly formulated by human and distinguished well by machine
Our attempt to improve things One Particular Realization The Signal Clustered signal Cluster distribution instances Interaction Resulting cluster distribution
Our attempt to improve things Implementation Details Reading signal in real time Signal Processing Same as was explained earlier & Exploring cluster distribution Exploration vs. Exploitation • Only explore – distribution will randomly change every time • Only exploit – there is a good Relaxed Imagine flower chance we missed best mental state and without exploration we will never find it Fear or frustration
Our attempt to improve things Results Current solution: K-means, e-greedy strategy Initial Distribution This is how it looks like after one minute of thinking about random things Relaxed Test subject was trying to be as relaxed as possible, not concentrating on anything particular Anxious Test subject was performing calculations in the mind
Conclusion Done • Skeleton solution which can be used to test strategies • Implementation of basic strategy • Learning a bit about nature of distributions in different mental states To Do • More careful signal pre-processing • Learn more about cluster consistency (longer trials, averaged results, different test subjects) • Find clustering algorithm which will fit nicely under multi-armed bandit formalization of exploitation-exploration trade-off (other than k-means) • Find best exploration-exploitation trade-off strategy for our type of data
Thank you! Please ask questions!
Recommend
More recommend
