BRAIN COMPUTER INTERFACE A SEMINAR REPORT

[seminar class]

[attachment=10052]
Brain Computer Interfaces
What is an EEG?
 An electroencephalogram is a measure of the brain's voltage fluctuations as detected from scalp electrodes.
 It is an approximation of the cumulative electrical activity of neurons.
What is it good for?
Neurofeedback
 treating ADHD
 guiding meditation
Brain Computer Interfaces
 People with little muscle control (i.e. not enough control for EMG or gaze tracking)
 People with ALS, spinal injuries
 High Precision
 Low bandwidth (bit rate)
EEG Background
 1875 - Richard Caton discovered electrical properties of exposed cerebral hemispheres of rabbits and monkeys.
 1924 - German Psychiatrist Hans Berger discovered alpha waves in humans and invented the term “electroencephalogram”
 1950s - Walter Grey Walter developed “EEG topography” - mapping electrical activity of the brain.
Physical Mechanisms
 EEGs require electrodes attached to the scalp with sticky gel
 Require physical connection to the machine
 Electrode Placement
 Standard “10-20 System”
 Spaced apart 10-20%
Letter for region
 F - Frontal Lobe
 T - Temporal Lobe
 C - Center
 O - Occipital Lobe
 Number for exact position
 Odd numbers - left
 Even numbers - right
 Electrode Placement
 A more detailed view:
 Brain “Features”
User must be able to control the output:
 use a feature of the continuous EEG output that the user can reliably modify (waves), or
 evoke an EEG response with an external stimulus (evoked potential)
 Continuous Brain Waves
 Generally grouped by frequency: (amplitudes are about 100µV max)
 Brain Waves Transformations
 wave-form averaging over several trials
 auto-adjustment with a known signal
 Fourier transforms to detect relative amplitude at different frequencies
 Alpha and Beta Waves
 Studied since 1920s
 Found in Parietal and Frontal Cortex
 Relaxed - Alpha has high amplitude
 Excited - Beta has high amplitude
 So, Relaxed -> Excited
means Alpha -> Beta
 Mu Waves
 Studied since 1930s
 Found in Motor Cortex
 Amplitude suppressed by Physical Movements, or intent to move physically
 (Wolpaw, et al 1991) trained subjects to control the mu rhythm by visualizing motor tasks to move a cursor up and down (1D)
 Mu Waves
 Mu and Beta Waves
 (Wolpaw and McFarland 2004) used a linear combination of Mu and Beta waves to control a 2D cursor.
 Weights were learned from the users in real time.
 Cursor moved every 50ms (20 Hz)
 92% “hit rate” in average 1.9 sec
 Mu and Beta Waves
 Movie!
 Mu and Beta Waves
 How do you handle more complex tasks?
 Finite Automata, such as this from (Millán et al, 2004)
 P300 (Evoked Potentials)
 occurs in response to a significant but low-probability event
 300 milliseconds after the onset of the target stimulus
 found in 1965 by (Sutton et al., 1965; Walter, 1965)
 focus specific
 P300 Experiments
 (Farwell and Donchin 1988)
 95% accuracy at 1 character per 26s
 P300 (Evoked Potentials)
 (Polikoff, et al 1995) allowed users to control a cursor by flashing control points in 4 different directions
 Each sample took 4 seconds
 Threw out samples masked by muscle movements (such as blinks)
 (Polikoff, et al 1995) Results
 50% accuracy at ~1/4 Hz
 80% accuracy at ~1/30 Hz
VEP - Visual Evoked Potential
 Detects changes in the visual cortex
 Similar in use to P300
 Close to the scalp
Model Generalization (time)
 EEG models so far haven’t adjusted to fit the changing nature of the user.
 (Curran et al 2004) have proposed using Adaptive Filtering algorithms to deal with this.
Model Generalization (users)
 Many manual adjustments still must be made for each person (such as EEG placement)
 Currently, users have to adapt to the system rather than the system adapting to the users.
 Current techniques learn a separate model for each user.
 (Müller 2004) applied typical machine learning techniques to reduce the need for training data.
 Support Vector Machines (SVM) and Regularized Linear Discriminant Analysis (RLDA)
 This is only the beginning of applying machine learning to BCIs!
BCI Examples - Communication
 Farwell and Donchin (1988) allowed the user to select a command by looking for P300 signals when the desired command flashed
BCI Examples - Prostheses
 (Wolpaw and McFarland 2004) allowed a user to move a cursor around a 2 dimensional screen
 (Millán, et al. 2004) allowed a user to move a robot around the room.
BCI Examples - Music
 1987 - Lusted and Knapp demonstrated an EEG controlling a music synthesizer in real time.
 Atau Tanaka (Stanford Center for Computer Research in Music and Acoustics) uses it in performances to switch synthesizer functions while generating sound using EMG.
In Review…
 Brain Computer Interfaces
 Allow those with poor muscle control to communicate and control physical devices
 High Precision (can be used reliably)
 Requires somewhat invasive sensors
 Requires extensive training (poor generalization)
 Low bandwidth (today 24 bits/minute, or at most 5 characters/minute)
Future Work
 Improving physical methods for gathering EEGs
 Improving generalization
 Improving knowledge of how to interpret waves (not just the “new phrenology”)
Invasive BCIs
 Have traditionally provided much finer control than non-invasive EEGs (no longer true?)
 May have ethical/practical issues
 (Chapin et al. 1999) trained rats to control a “robot arm” to fetch water
 (Wessberg et al. 2000) allowed primates to accurately control a robot arm in 3 dimensions in real time.