BrainComputer Interfaces Current Research and Future Prospects

1
Brain-Computer Interfaces Current Research and
Future Prospects
2
What is a Brain-Computer Interface (BCI)?

• Traditional Human-Computer Interaction (HCI)
  Human controls virtual or physical objects using
  muscular activity. Examples
• Control of a cursor using a mouse (hand/finger
  movements)
• Typing using a keyboard (finger/hand movements)
• Remote control of robotic devices using a
  joystick (hand/arm movements)
• Control of an automobile using a steering wheel
  and pedals (hand/arm/feet/leg movements)
• Brain-Computer Interface (BCI) A device that
  utilizes brain activity for direct control of
  physical or virtual objects without relying on
  muscular activity or body movements.

3
Why BCI?

• Improved communication and control for paralyzed
  and locked-in patients (e.g. stroke, ALS, spinal
  injury patients)
• Applications in health and safety
• E.g. Early detection and diagnosis of symptoms
• E.g. Alertness monitoring in critical occupations
  (e.g. night drivers, pilots, railway engineers)
• Computer-aided education and learning
• E.g. Brain-activity based presentation of
  material?
• Augmented cognition (brain-body actuated control)
• E.g. Air Force research using hybrid brain-body
  interfaces for speeding up responses during
  flight
• Entertainment and Security
• E.g. Video games, TV/web browsing for patients,
• E.g. Better lie detection devices and brain
  fingerprinting?

4
BCI What is involved?
From (Nicolelis, 2001)
5
What is EEG?
Scalp electrode

• Voltage fluctuations at the scalp due to
  activities of large populations of neurons in the
  cerebral cortex
• Input potentials and activities of neurons get
  attenuated and summated due to passage through
  meninges, cerebrospinal fluid, skull, and scalp.

EEG
Electrical activity
Pyramidal neurons in cerebral cortex
6
Where do I get some?

• EEG signals Acquired from a cap of electrodes
  that contact scalp through a gel
• Recent progress Active electrodes and dry
  electrodes.
• Signals are in microvolts range ? need to be
  amplified

10-20 arrangement of scalp electrodes
7
Does it mean anything?
7.5-13 Hz
Alpha waves Associated with unfocusing attention
(relaxation)
Mu waves Associated with movements or intention
to move
lt 3 Hz
gt 14 Hz
Delta waves Associated with deep sleep
Beta waves Associated with alertness and
heightened mental activity
(Images from Scientific American, 1996)
8
Does it mean anything? (2)
P300 A characteristic EEG waveform elicited
about 300 milliseconds after spotting a
relatively rare stimulus (e.g. spotting the store
you were looking for while driving)
No P300
9
Butwhat good is it?

• Typing words by flashing letters and detecting
  P300s (Farwell Donchin, 1988)
• Select a character (out of 36) in 26 seconds with
  95 accuracy
• Move a cursor towards a target on a screen by
  training subjects to control the amplitude of
  their Mu waves (Wolpaw et al., 1991
  Pfurtscheller et al., 1993)
• 10-29 hits/min and 80-95 accuracy after 12
  45-min sessions
• Moving a joystick in 1 of 4 directions by
  recognizing readiness potentials and
  classifying EEG patterns during mental tasks
  using artificial neural networks (Hiraiwa et al.,
  1993 Anderson Sijercic, 1996)

10
The Fine Print

• Good temporal resolution but poor spatial
  resolution
• Poor quality signals
• Considerable attenuation and background noise
• Serious issues with artifacts (eyeblink, head/eye
  movement)
• Variability across subjects, sessions (e.g.,
  electrode positioning)
• Attention wandering, motivation of subjects
• Low SNR how much information is actually
  extractable?
• Moving target subjects are also adaptive
• What feedback/adaptation should system use?
• Unintuitive/cumbersome communication paradigms

11
Invasive BCIs Theres hope elsewhere
Extracellular recording of neural spikes
(From Shenoy et al., 1999)
Array of silicon electrodes with platinum-plated
tips
Array is implanted in an area of the cerebral
cortex
12
Look Ma, No Hands! Neural Control by a Rat
Electrode Array
Water (Reward)
Robot Arm
Lever
Switch to select between BCI/Lever Control
Recorded activities of 24 motor cortex neurons
Neural Population Function
Spikes from 2 motor cortex neurons
(Chapin et al., 1999)
13
Neural Control of a Robotic Arm by a Rat
Experiment by Chapin et al., 1999

• Rat presses a lever to move a robotic arm to get
  reward
• Neural outputs from rats motor cortex train an
  artificial
• neural network to control the robotic arm
• After training, several rats no longer used
  their own body
• movements but retrieved reward using their
  neural activity

14
Our Goals

• Build generative models for EEG
• Try to filter out noise, variability and
  artifacts
• Model the attenuation and spatial filtering of
  scalp
• Independent preprocessing step
• Propose and evaluate more powerful experimental
  paradigms
• More natural schemes e.g., joystick based
  cursor control
• Higher bit rate
• Adaptive and online algorithms
• Explore the rangeneural gt ECoG gt EEG
• With collaborators from the med school/primate
  research center

15
Finally! Some details

• Proposal
• Ê(t1) f(Ê(t)) m (Ê is true EEG)
• E(t) g(Ê(t)) n (E is observed EEG)
• Hypothesize scalp EEG as being generated by a
  noisy dynamical system.
• Model the internal dynamics, observation
  process, and noise
• Use model to filter EEG before use in generic BCI
  application
• Issues
• What is the true EEG? A case of unsupervised
  learning.
• What model to use? First cut Linear dynamics,
  gaussian noise.
• How to evaluate model and its genericity?

16
A Pilot Study Semisupervised learning

• Ê(t1) A Ê(t) m (Linear Dynamics)
• E(t) M Ê(t) n (M ICA mixing matrix)
• A captures linear dynamics
• M is an estimated ICA mixing matrix
• Paradigm
• Mix 3 signals and noise to form the vector E
• Take portion of dataset for training
• Estimate mixing matrix M, generate supervisory
  signal Ê
• Estimate dynamics and noise parameters
• Take portion of dataset for testing
• Filter using learned dynamical system

17
Sample Graphs
18
Sample Graphs
19
(No Transcript)
20
(No Transcript)
21
(No Transcript)
22
Claims and Caveats

• Clearly a toy example to illustrate the method
• However, useful if dynamics are well-captured
• What about EEG?
• More complex dynamics
• Variable behavior over time hierarchical
  generative models?
• ICA/PCA need not necessarily work