Cognitive Control Signals for Neural Prosthetics

1
Cognitive Control Signals for Neural Prosthetics

• Sam Musallam
• The Andersen Lab

2
Cognitive-based rather thanmotor-based

• May require fewer cells, less invasive
• Cognitive variables-expected value
• Hierarchical control

PMC
3
We are currently implanting 160 electrodes in
PRR and in DPMC
- 32 electrode arrays - Made from platinum and
iridium - 80 microns at the shaft - 2-3 micron
tip
4
Arrays are Implanted in the Medial Intraparietal
Sulcus and Dorsal Premotor Cortex
5
Arrays are Implanted in Medial Intraparietal
Sulcus
6
Decoding Goals E
Reach Trials
Decode Trials
7
Reach Trials
8
Neuron example from parietal cortex
9
Decoding Goals E
Reach Trials
Decode Trials Bayes rule
10
Brain Control Trials
11
Example of feedback performanceParietal
Fixation maintained throughout trial.
12
Brain control trials using neurons from DPMC
75
13
Frozen or adaptive database? Doesnt seem to
matter
14
At 100 ms, predictions are still significantly
above chance
15
Overall performance improves with time
Mutual Information
Monkey Performance
16
Identical but different trials can elicit
different firing rates
17
In a single session, feedback performance
fluctuates
6 Targets
Chance
18
Type of reward changes the tuning of neurons in
PRR
Blackorange juice Red water
19
Tuning is also enhanced for increased probability
and magnitude of reward
20
Our decoding ability improves with preferred
reward condition
21
MI increases with preferred reward trials
22
Overall decode improvement
23
Decoding the value of reward
24
Decoding Direction and Reward
25
More neurons improve the decode
26
This behavior is advantageous for prosthetic
control

• Optimize control of prosthetic devices under
  given decoding constraints
• counter neuronal sample biases(surgical
  placement of electrodes, etc.)
• counter signal non-stationarities(changes in
  tissue, electrode characteristics, etc.)
• allow multiple tool use(update or variable
  functionality of prosthesis)

27
What about trajectories?
28
80 accuracy to 8 targets
Mean time taken to reach target 700
ms. Fixation point is randomized.
29
Conclusion

• Cognitive neural activity not directly related to
  visual input or motor output can be used for as
  prosthetic control signals.
• Cognitive signals can also give us information
  about the patients preference or mood
• Signals in higher level areas like the PPC and
  DPMC can be used to guide trajectories as well

30

• Brian Corneil
• Hans Scherberger
• Bradley Greger
• Grant Mulliken

31
Thank You
32
(No Transcript)