2. http://nrg.mbi.ufl.edu
Enabling Neurotechnologies for Overcoming
Paralysis
Develop direct neural
interfaces to bypass
injury. Communicate
and control (closed-
loop, real-time) directly
via the interface.
Leuthardt
4. http://nrg.mbi.ufl.edu
What are the Building Blocks?
Signal Sensing
Amplification
Pre-Processing
Telemetry
Interpret Neural
Activity
Control
Scheme
Feedback
Closed Loop BMI
Provide neurophysiologic basis and engineering theory for a fully implantable neural
Interface for restoring communication and control
5. http://nrg.mbi.ufl.edu
BMI lessons learned
Relationship between user
and BMI is inherently
lopsided. Users are
intelligent and can use
dynamic brain organization
and specialization while
BMIs are passive devices
that enact commands
I/O models have difficulty
contending with new
environments without
retraining
Laboratory BMIs need to be
better prepared for ADL
7. http://nrg.mbi.ufl.edu
Vision for Next Generation Brain-
Machine Interaction
Intelligent behavior arises from
the actions of an individual
seeking to maximize received
reward in a complex and
changing world.
Perception-Action Cycle:
Adaptive, continuous process
of using sensory information to
guide a series of goal-directed
actions.
10. http://nrg.mbi.ufl.edu
Co-Adaptive BMI involves TWO intelligent
agents involved in a continuous dialogue!!!
ROBOT
actions
rewards
brainstates
RAT’S BRAIN
environment
RAT’S BRAIN
COMPUTER AGENT
11. http://nrg.mbi.ufl.edu
Decoding using
Reinforcement Learning
Rather than knowledge of the kinematic hand trajectory
only a performance score is supplied. The score could
represent reward or penalty, but does not directly provide
information about how to correct for the error.
Reward based learning - try to choose strategy to
maximize rewards.
RL originated from optimal control theory in Markov
Decision Processes.
Rt = γn−t+1
rn
n= t+1
∞
∑
Q st ,at( )
*
= E Rt | st ,at{ }
12. http://nrg.mbi.ufl.edu
Experimental Co-Adaptive BMI Paradigm
-3 -2 -1 0 1 2 3
0
1
2
3
0
1
2
3
Incorrect
Target
Correct
Target
Starting
Position
Match LEDs
Grid-space
Match LEDs
Rat’s Perspective
Water Reward
Map workspace
to grid
Rat
Robot Arm
Left Lever Right Lever
27 discrete actions
26 movements
1 stationary
13. http://nrg.mbi.ufl.edu
Agent - Value function estimation
Qk (
v
st ) = tanh si,t
i
∑ wij
⎛
⎝
⎜⎜
⎞
⎠
⎟⎟
⎛
⎝
⎜
⎜
⎞
⎠
⎟
⎟w jk
j
∑ δt = rt +1 +γQ(st+1,at +1) −Q(st ,at )
15. http://nrg.mbi.ufl.edu
Key Concepts for the Future
Fully implantable interfaces are only half of
the story.
Sharing of goals enables brain-computer
dialogue and symbiosis
Need for intelligent decoders that assist and
co-adapt with the user.
16. http://nrg.mbi.ufl.edu
History of Man-Machine
Interaction
“Implanting tiny machines
into the nerves of the heart
would make us less human”
Today, over half a million
pacemakers are implanted
annually!
We are at the frontier for
integrating machines with
the nervous system to
restore and enhance
function.
Nicolelis
What are we going to do with this information? Bring together a novel BMI architecture.
Active assistant to the paralyzed patient.
Interaction is not passive but grows through experience.
The hidden layer of the MLP is a projection space where the state signal can be segmented. 3 hidden PEs was chosen based on some offline analysis. From this we can use a temporal difference error based ONLY on the rewards and our own predictions. Again details can be discussed offline – otherwise it is too boring.
Action 3 example