Wireless neural recording systems are subject to stringent power consumption constraints to support long-term recordings and to allow for implantation inside the brain. In this paper, we propose using a combination of on-chip detection of action potentials (“spikes”) and compressive sensing (CS) techniques to reduce the power consumption of the neural recording system by reducing the power required for wireless transmission. We empirically verify that spikes are compressible in the wavelet domain and show that spikes from different neurons acquired from the same electrode have subtly different sparsity patterns or supports. We exploit the latter fact to further enhance the sparsity by incorporating a union of these supports learned over time into the spike recovery procedure. We show, using extracellular recordings from human subjects, that this mechanism improves the SNDR of the recovered spikes over conventional basis pursuit recovery by up to 9.5 dB (6 dB mean) for the same number of CS measurements. Though the compression ratio in our system is contingent on the spike rate at the electrode, for the datasets considered here, the mean ratio achieved for 20-dB SNDR recovery is improved from 26:1 to 43:1 using the learned union of supports. http://nesl.ee.ucla.edu/document/show/364

1. Compressively Sensing
Action Potentials
(Neural Spikes)
Zainul Charbiwala, Vaibhav Karkare, Sarah Gibson
Dejan Marković, Mani B. Srivastava
In collaboration with UCLA Neuroscience

2. Full Implantability
Cyberkinetics Neurotechnology Systems/Matthew McKee Utah Electrode Array
zainul@ee.ucla.edu - Spike CS - May 2011 2

3. Full Implantability
Put the exam room IN the patient !
Cyberkinetics Neurotechnology Systems/Matthew McKee Utah Electrode Array
zainul@ee.ucla.edu - Spike CS - May 2011 2

4. Raw Signal from a Depth Electrode
[uwhealth.org]
zainul@ee.ucla.edu - Spike CS - May 2011 3

5. Signal Band Pass Filtered to Enhance Spikes
Filtering
zainul@ee.ucla.edu - Spike CS - May 2011 4

6. Spike Detection
Detection
These spikes are from
multiple neurons in the
vicinity of the electrode
zainul@ee.ucla.edu - Spike CS - May 2011 5

7. Spike Alignment
Alignment
zainul@ee.ucla.edu - Spike CS - May 2011 6

8. Spike Sorting
Sorting
zainul@ee.ucla.edu - Spike CS - May 2011 7

9. Traditional Wireless Neural Recording Process
In vivo
Amplify Band Pass Spike Detect Radio
and Digitize Filter and Align TX
Radio Spike
RX Sorting
Ex vivo
zainul@ee.ucla.edu - Spike CS - May 2011 8

10. CS Neural Recording Process
In vivo
Amplify Band Pass Spike Detect Compressive Radio
and Digitize Filter and Align Sampling TX
Radio CS Spike
RX Recovery Sorting
Add
Support
Ex vivo
zainul@ee.ucla.edu - Spike CS - May 2011 9

11. Two Features Useful for Compressed Sensing
‣ Spikes are compressible in the wavelet domain
‣ Spikes from the same neuron are similar in morphology
zainul@ee.ucla.edu - Spike CS - May 2011 10

12. Representing the Spike in DWT Domain
Few large coefficients in
DWT domain --
compressible
zainul@ee.ucla.edu - Spike CS - May 2011 11

13. Number of Significant Coefficients (Support) in
DWT Domain
zainul@ee.ucla.edu - Spike CS - May 2011 12

14. Sparsity Inducing Transform
x
n
zainul@ee.ucla.edu - Spike CS - May 2011 13

15. Sparsity Inducing Transform
x
n
DWT - Ψ
zainul@ee.ucla.edu - Spike CS - May 2011 13

16. Sparsity Inducing Transform
x
n
DWT - Ψ
z= Ψ x
n
zainul@ee.ucla.edu - Spike CS - May 2011 13

17. A Visual Tour of Compressed Sensing
x
n
zainul@ee.ucla.edu - Spike CS - May 2011 14

18. A Visual Tour of Compressed Sensing
x
n
m Φ
zainul@ee.ucla.edu - Spike CS - May 2011 14

19. A Visual Tour of Compressed Sensing
x
n
m Φ
y= Φ x
m
zainul@ee.ucla.edu - Spike CS - May 2011 14

20. Compressed Sensing Recovery
y
m
zainul@ee.ucla.edu - Spike CS - May 2011 15

21. size of the support computed over 600 000 spikes extracted
from the entire dataset. About Recovery describe the action
Compressed Sensing 8 coefficients
potential adequately, a compression of 6:1 from the raw 48
samples acquired per spike.
y
We can now formulate our recovery procedure as the basis
m
pursuit denoising (BPDN) [5] problem:
1 −1 2
z = argmin y − ΦΨ z
ˆ ˜ 2
+λ z
˜ 1 (2)
z 2
˜
zainul@ee.ucla.edu - Spike CS - May 2011 15

22. size of the support computed over 600 000 spikes extracted
from the entire dataset. About Recovery describe the action
Compressed Sensing 8 coefficients
potential adequately, a compression of 6:1 from the raw 48
samples acquired per spike.
y
We can now formulate our recovery procedure as the basis
m
pursuit denoising (BPDN) [5] problem:
1 −1 2
z = argmin y − ΦΨ z
ˆ ˜ 2
+λ z
˜ 1 (2)
z 2
˜
−1
ˆ ˆ
x=Ψ z
n
zainul@ee.ucla.edu - Spike CS - May 2011 15

23. CS Neural Recording Process
In vivo
Amplify Band Pass Spike Detect Compressive Radio
and Digitize Filter and Align Sampling TX
Radio CS Spike
RX Recovery Sorting
Ex vivo
zainul@ee.ucla.edu - Spike CS - May 2011 16

24. Conventional Basis Pursuit Results
3rd Quartile
Median over 0.6M spikes
1st Quartile
35.7 measurements for
20dB SNDR
18x → 26x compression
zainul@ee.ucla.edu - Spike CS - May 2011 17

25. Two Features Useful for Compressed Sensing
‣ Spikes are compressible in the wavelet domain
‣ Spikes from the same neuron are similar in morphology
zainul@ee.ucla.edu - Spike CS - May 2011 18

26. Similarity in Neural Spikes
Little support mismatch,
Mismatch at low values
zainul@ee.ucla.edu - Spike CS - May 2011 19

27. potential adequately, a compression of 6:1 from the raw 48
samples acquired per spike.
We can now formulate our recovery Recovery the basis
Incorporating Similarity in CS procedure as
pursuit denoising (BPDN) [5] problem:
Basis Pursuit Recovery
1 −1 2
z = argmin y − ΦΨ z
ˆ ˜ 2
+λ z
˜ 1 (2)
z 2
˜
Φ - Sampling matrix
Ψ - Sparsifying transform
λ - L1 penalizing factor
zainul@ee.ucla.edu - Spike CS - May 2011 20

28. potential adequately, aΦΨ−1 satisfies a condition known a
the ensemble matrix compression of 6:1 from the raw 48
samples acquired per spike.
the restricted isometry property (RIP) [5], the error in th
Incorporating SimilaritywillCS Recovery the basis
formulate our in be procedure as
We can nowabove problemrecoverystable and bounded wit
solution to the
pursuit denoising (BPDN) [5] problem:
overwhelming probability.
Basis Pursuit Recovery
1
In [7], Lu=and Vaswani−introduced + λnew approach t
−1 2 a
z argmin y ΦΨ z 2
ˆ ˜ z 1
˜ (2)
z 2
BPDN called Modified-CS when additional knowledge i
˜
available. Specifically, they show that if the support of th
Φ - Sampling matrix
spike -waveform (or a part thereof) was known a priori, th
Ψ Sparsifying transform
λ - L1 penalizing factor
error in the solution to Eq. (2) admits a lower bound. Thei
modified BPDN approach is given by:
Masked Basis Pursuit Recovery
1 −1 2
z = argmin y − ΦΨ z 2 + λ zT c 1
ˆ ˜ ˜ (3
z 2
˜
c
where, T is the complement of the known support so zT ˜
denotes the elements in z that are not included within T . Th
˜
bound on the 2 norm of the solution error depends on λ 20 an
zainul@ee.ucla.edu - Spike CS - May 2011

29. potential adequately, aΦΨ−1 satisfies a condition known a
the ensemble matrix compression of 6:1 from the raw 48
samples acquired per spike.
the restricted isometry property (RIP) [5], the error in th
Incorporating SimilaritywillCS Recovery the basis
formulate our in be procedure as
We can nowabove problemrecoverystable and bounded wit
solution to the
pursuit denoising (BPDN) [5] problem:
overwhelming probability.
Basis Pursuit Recovery
1
In [7], Lu=and Vaswani−introduced + λnew approach t
−1 2 a
z argmin y ΦΨ z 2
ˆ ˜ z 1
˜ (2)
z 2
BPDN called Modified-CS when additional knowledge i
˜
available. Specifically, they show that if the support of th
Φ - Sampling matrix
spike -waveform (or a part thereof) was known a priori, th
Ψ Sparsifying transform
λ - L1 penalizing factor
error in the solution to Eq. (2) admits a lower bound. Thei
modified BPDN approach is given by:
Masked Basis Pursuit Recovery
1 −1 2
z = argmin y − ΦΨ z 2 + λ zT c 1
ˆ ˜ ˜ (3
z 2
˜
T - Set of support indices expected in the spike known support so zT
c
where, T is the complement of the ˜
Tc - Set of support indices not expected ­ sparser than support itself
denotes the elements in z that are not included within T . Th
˜ [Wang, et. al., 2010; Vaswani et. al., 2010, Charbiwala, 2009]
bound on the 2 norm of the solution error depends on λ 20
zainul@ee.ucla.edu - Spike CS - May 2011 an

30. Learning a Union of Supports Results in Very
Low Sparsity
zainul@ee.ucla.edu - Spike CS - May 2011 21

31. Learning a Union of Supports Results in Very
Low Sparsity
zainul@ee.ucla.edu - Spike CS - May 2011 21

32. Learning a Union of Supports Results in Very
Low Sparsity
zainul@ee.ucla.edu - Spike CS - May 2011 21

33. CS Neural Recording Process
In vivo
Amplify Band Pass Spike Detect Compressive Radio
and Digitize Filter and Align Sampling TX
Radio CS Spike
RX Recovery Sorting
Add
Support
Ex vivo
zainul@ee.ucla.edu - Spike CS - May 2011 22

34. Learned Union of Support Results
22.1 measurements for
20dB SNDR
18x → 43x compression
zainul@ee.ucla.edu - Spike CS - May 2011 23

35. Comparison with Oracle Support Knowledge
zainul@ee.ucla.edu - Spike CS - May 2011 24

36. Stability of Union Progression
zainul@ee.ucla.edu - Spike CS - May 2011 25

37. 8 25
Size of Union of Supports
7 20
Why Union CS Works
6 15
Error bound
5 10
4 I nc re asi ng |∆|, Fi x e d |∆ e| 5
I nc re asi ng |∆|, D e c re asi ng |∆ e|
3 0
0 100 200 300 400 500 600
Fi x e d |∆|, I nc re asi ng |∆ e|
2 #Spikes
D e c re asi ng |∆|, I nc re asi ng |∆ e| Fig. 5. Median progression of the size of the learn
1 each set of 1000 spikes.
2 4 6 8 10 12 14 16 18 20
Size of support set 4
x 10
2.5
Fig. 4. Trend lines of the error bound that trade off the size of the unknown
Norm Outside
#Spike Occurences
support, ∆ with the size of the superfluous support, ∆e . The curves illustrate 2 Norm Outside
the sensitivity of the error bound on |∆| and the relative insensitivity to |∆e |. µ = 0 . 036, σ= 0. 037
1.5
We provide two justifications for our union of supports
1 µ = 0. 2 72, σ= 0. 167
proposal, one derived analytically and one empirically from
Δ : Number ofthe first, Negativesthe Unknown Support 0.5
our datasets. For False we excerpt -- bound on modified
Δe : Number of False Positives -- Superfluous Support 00 Adding to superfluous set is 0.6
BPDN reconstruction error from [7]:
0.1 0.2 0.3 0.4 0.5
2
θ|T |,|∆| alright if unknown support Norm
1
z − z 2 ≤ λ |∆|
ˆ 2 +1 2
Fig. 6. set is reducing norm of signal outside
Histogram of the at same pace
θ|T |,|∆|
1 − δ|T | 1 − δ|∆| − 1−δ|T | and outside learned union of support of all preced
σ 2 sizes of the two sets. We defer a full an
+ (4) of the error bound performance for futur
1 − δ|N ∪∆e |
Fig. 5 shows the progression of the size
where |·| refers to set cardinality, δs is the s-restricted isometry 2011 T over our datasets. In order to co
set
zainul@ee.ucla.edu - Spike CS - May
constant constant [5] for the matrix ensemble ΦΨ−1 and θ
26

38. Spike Sorting Performance
U. Rutishauser, , A. Mamelak, and E. Schuman, “Online detection and sorting of extracellularly recorded action potentials in human
medial temporal lobe recordings, in vivo,” Journal of Neuroscience Methods, vol. 154, pp. 204–224, 2006.
zainul@ee.ucla.edu - Spike CS - May 2011 27

39. Clustering Performance
Original CS Neural Recording
16 measurements for
90% clustering accuracy
18x → 56x compression
zainul@ee.ucla.edu - Spike CS - May 2011 28

40. Power Consumption Comparison
zainul@ee.ucla.edu - Spike CS - May 2011 29

41. Conclusions and Future Directions
‣ 2x compression versus sensing raw spikes at low power
‣ Sorting possible on compressed version at lower rates
‣ In development: 8-channel FPGA based implementation
‣ Within the year: ASIC based 16-channel solution
zainul@ee.ucla.edu - Spike CS - May 2011 30

42. Thank You !
Slides, Code, Data @
http://nesl.ee.ucla.edu/people/zainul

43. Backup Slides

44. Magnitude of Coefficients of Consecutive
Spikes are Dissimilar
Spikes differ by 90%
on the average in DWT
domain
zainul@ee.ucla.edu - Spike CS - May 2011 33

45. Magnitude of Mismatched Coefficients in
Consecutive Spikes is Lower
Spikes are 70% sparser
outside support of
previous spike
zainul@ee.ucla.edu - Spike CS - May 2011 34

46. Magnitude of Mismatched Coefficients from
All Previous Spikes is Very Low
Spikes are 95% sparser
outside support of all
previous spikes
zainul@ee.ucla.edu - Spike CS - May 2011 35

47. Comparison with Modified-CS
zainul@ee.ucla.edu - Spike CS - May 2011 36

48. Sorting Performance with Modified-CS
zainul@ee.ucla.edu - Spike CS - May 2011 37