-
[Picture courtesy of K. Diesseroth, Stanford]
-
An Exciting Time for Neuroscience
-
Brain-Machine Interfaces (BMI) Making News
Listening to the voices inside your head “Neuroscientists may one day be able to hear the imagined speech of a patient unable to speak due to stroke or paralysis, according to University of California, Berkeley researchers.” [Pasley at al, PLOS12]
-
Some Interesting Numbers
Human Brain: 80 billion neurons, 100 trillion synapses 1200 cm3, 1300 gram 15-20 W, 1017 floating point ops/sec
Mouse Brain: 75 million neurons, 100 billion synapses 420 mm3, 0.5 gram 15 mW
Sampling every neuron in a mouse brain at 1 bit resolution at 1 kHz: 75 Gbit/sec
-
Instrumenting the Brain
[Courtesy: Marblestone at el, 2013]
-
Imaging the Brain for BMI
[Schwartz et al. Neuron, 2006]
-
BMI Interface Challenges: Acquire and Transmit Neural Signals Reliably and Consistently Over Long Lifespans (> 10 Years)
Weak signals, Low SNR, High Offset, Mixture of various signals (field and action potentials)
ADC LNA
electrodes
DSP
memory
Tx
regulator
clock
1 2 3 4 5−200
−150
−100
−50
0
50
100
150
200
Time (s)
Inpu
t Vol
tage
(μV)
Full WaveformDelta Band
ECoG
-
The Berkeley Brain-Machine Interface (BMI) - Combining “Neural Dust” and μECoG
“An implanted neural interface that can provide imaging (and possibly stimulation) of neural activity at multiple scales of resolution using arrays of patterned and free-floating sensors”
A collaborative effort between BWRC, UCB Engineering, UCB Neuroscience, and UCSF NeuroSurgery (as part of the CNEP Center)
Key requirement: reliability and longevity Tons of selectable channels Wireless powering and data
communications Compliant, flexible and
microscopic (!)
-
Physical Interface Platforms across Scale and Modality
μECoG+BMIPeter Ledochowitsch / Aaron Koralek
Carmena / Maharbiz
-
Example: High-density Nanotrodes
• Optical excitation, electrical readout
• Ultra-compliant parylene cables allow truly floating electrodes
• Scalable and manufacturable
• Ultra-high density electrical recording (64-128 channels)
• 16 integrated 3 µm thick optical waveguides on each probe
-
A Library of Acquisition, Communication and Power Delivery Components!
2.4mm
64 channel remotely powered wireless uECoG [Muller, Le, Ledoschowicz, Li]
Free-floating wireless AP acquisition electrodes [Biederman, Yeager, VLSI12]
Register Bank
Memory Feature Extraction
Preamble Buffer
Spike Detection
Spike Alignment
Asynchronous 250 nW/channel spike-sorting [Liu, VLSI12]
-
The Neural Sensor Node: Size, Power/Energy and Bandwidth
Wireless powering only realistic option But limited by safety concerns
Bandwidth requirements set by # channels and information resolution
-
Reference: AP Acquisition Channel
[R. Muller et al, ISSCC 2011]
0.013mm2, 5uW DC Coupled Neural Signal Acquisition IC with 0.5V Supply (65 nm CMOS) Digital transistors are
cheap in advanced CMOS processes
Avoid off-chip components by dc-coupling and filtering in digital domain
No high-precision analog components
-
Getting to the Point – Some System Examples
-
Wireless μECoG may provide up to 1000 channels with pitch as low as 200 μm.
Antenna + electrodes printed on parylene substrate using semiconductor-like process
Providing unprecedented resolution and offering huge potential for BMI (ALS, Epilepsy).
[Courtesy: P. Ledocowich, R. Muller, M. Maharbiz, J. Rabaey]
Circuit elements similar to AP sensor nodes
A Global View with μECoG
64 channel array
-
Loop Antenna for Link Optimization
0 200 400 600 800-20-19-18-17-16-15
f [MHz]
MA
G [
dB
]
0 200 400 600 8000
100200300400
f [MHz]
Ptx
[m
W]
0 200 400 600 8000123456
f [MHz]
Prx
[m
W]
[Courtesy: T. Bjorninen, R. Muller]
Power availability limits functionality
-
[Courtesy: R. Muller, W. Li, H. Le, S. Gambini]]
65 nm CMOS 2.4 uW/channel @ 0.5V 0.025 mm2/channel 1 Mbits/sec data rate
Confidential – pre-publication A 240 uW Self-Contained Wireless uECoG System
-
Neural Dust v1 – Free-floating Electrodes
μ
[Biederman, Yeager, VLSI12]
Measured Power
Layout Area (um2)
4 Neural Amplifiers
8 uW 100 x 450
Total System: 10 uW 220 x 450
-
Node Architecture
4x Φ Sample
10b ADC LNA VGA DAC
-
Verification of Wireless Power Transfer
Slide 20
0
5
10
15
20
25
0 0.5 1 1.5 Pat
h Lo
ss (d
B) /
Pow
er (d
Bm
)
Transmission Distance (mm)
TX Power in Air (Meas)
Path loss in Air (Sim)
Edge-edge channel loss was verified using a micromanipulator in air and compared to simulation
Simulated channel loss in brain tissue is ~6db higher
Matching Network
IC
SMA
TX Coil 3mm x 3mm
-
Multi-Node Communication Programmable Miller subcarrier
1-10 MHz 6 simultaneous sensors
-
Neural Dust v2 – The Network Emerges
Skull
Brain
Interrogator
Power / Comm. Transcranial Link
Radio
Utilize “chiplet” designs to create dynamic system building blocks for a scalable & reconfigurable system
FlexSubs
Power & Data Controller
Can record or stimulate from each electrode & monitor electrode impedance
Scalable power output and data aggregation
Amp + ADC chiplet
Tether free data link to signal processing unit
Compliant Tether Improved scalability & power
transfer efficiency
-
Neural Dust v∞ - Scaling to Thousands
Electro-magnetic data communication and powering becomes extremely inefficient when nodes become very small (< 100 μm)
Yet optimal node size is around 10 μm
The ultra-sound opportunity Wavelength of 1MHz Ultrasound wave in brain = 1.562mm Wavelength of 1.5GHz RF wave in brain = 20 mm
Better resolution, better power-transfer efficiency
[DJ Seo et al, White Paper, June 2103]
-
The Data Deluge
Data link throughput limited by bandwidth and power constraints Passive RFID-like radio: 1-2 Mbit/sec @ 10 pJ/bit Active radio could do considerably more (≥ 10
Mbit/sec), but consumes more power/bit (> 100 pJ/bit)
Acquisition power/channel = 2-5 μW (amplification, conversion)
Action Potential Channel ECoG Channel 20 kHz x 10 bits = 200 kBit/sec 1 KHz x 12 bits = 12 kBit/sec
The challenge: How to scale to larger channel counts (1000’s or millions?)
-
Trade-off in Rate-Limited Systems
Low Channel Count High Channel Count
Full Information
Little Information
Current Systems
Our Goal
Goals and expectations differ: • Neuroscience requires full waveforms • Brain-machine interfaces operate from firing rates
(AP) or energy binning (ECoG, LFP)
-
Action Potentials
[Karkare, JSSC’11]
Straightforward compression yields little reduction (< 2) Better approach:
exploit sparseness – spike detection and clustering
Data Rates (including overhead) Raw 213 kBit/sec 5 channels Spike timing + epochs 51 kBit/sec 20 channels Firing rates 160 Bit/sec 6000 channels
(assuming 100 Hz avg. firing rate)*
-
Spike Detection and Feature Extraction
Example: NEO + Max & Epochs
[Courtesy: D. Markovic, UCLA; N. Narevsky, UCB]
-
The Energy Cost
Register Bank
Memory Feature Extraction
Preamble Buffer
Spike Detection
Spike Alignment
• Fully asynchronous to minimize leakage
• 250 nW/channel @ 0.25V • 0.03mm2 in 65nm CMOS
[Liu, VLSI 2012]
Spike detection + feature extraction
V. Karkare et al., A-
2 uW and 0.07 mm2 / channel [Karkare, Markovic, 2009]
Energy cost equal or smaller then acquisition
-
Clustering and Sorting One electrode may observe 1-10 neurons (or even more)
Commercial spike sorting software
V. Karkare et al.,
On-line spike sorting [Karkare, 2011]
Clustering algorithms
[Courtesy: D. Markovic, UCLA]
-
If lossy data compression is allowed - possible to transmit 1000’s of channels over single data link
The challenge resides now in the power domain: How to simplify acquisition front-end? (information extraction in the analog domain, compressive sampling, )
-
DC Offset ±50mV
ECoG 1Hz-500Hz 1µV-100µV
volta
ge
time
1 2
1 4 7 15 30 65 250
EEG EcOG Only
Anesthesia Sleep
Arousal Drowsiness Relaxed
Eye Closing
Alert/ Active Activity
Sensorimotor
-
ECoG (LFP, EEG)
[Heldman, Tran. Neur. Sys. 2006]
• Little information to be gained from time domain
• Most relevant: deviation from baseline in frequency bins
• Again: straightforward compression not effective
-
There is Hope … Compressed sampling
• Compression between 2 and 24 • Feature extraction in compressed
domain • But still needs full acquisition
[Shoaib, Verma, 2012]
-
Compressed Sampling in the Analog Domain?
Recovered spectrum [Courtesy: F. Maksimovic]
-
Exploit Spatial Correlation and/or Activity Sparseness?
0 0.02 0.04 0.06 0.08 0.1 0.12-0.1
0
0.1
Time [s]
Vol
tage
[m
V]
0 0.02 0.04 0.06 0.08 0.1 0.12-0.1
0
0.1
Vol
tage
[m
V]
Time [s]
0 0.02 0.04 0.06 0.08 0.1 0.12-0.1
0
0.1
Time [s]
Vol
tage
[m
V]
0 10 20 30 40 50 60 70 80
10
20
30
40
50
60
Frequency [Hz]
Ele
ctro
de #
1 2 3 4 5 6 7 81
2
3
4
5
6
7
8Image plot of all electrode voltages at a single time
-0.31
-0.3
-0.29
-0.28
-0.27
-0.26
-0.25
-0.24
-0.23
-0.22
-0.21
But time-domain image is not sparse
-
Plenty of Room for Creativity in the Million Neuron March
Innovation in wireless transceivers and
Neural swarm networking
Beam-forming Adaptation Learning
-
Final Reflections …. Brain-Machine Interfaces the ultimate in immersive
technologies - The potential is huge - Societal impact first, human advancement next
ULP circuit and systems design in concert with innovative technologies to provide “cellular electronics”
- Requires careful balance between energy availability, data bandwidth and information contents
Signal processing, interpretation in conjunction with control to play an important role
A new breed of integrated mixed
bio/electronics devices emerging - BioCyber (biotic-abiotic) systems a
formidable bet for EECS!
-
Acknowledgements: The many contributions of Elad Alon, Jose Carmena, Edward Chang, Bob Knight, Michel Maharbiz, K. Ganguly, Leena Ukkonen, Bruno Olshausen, Dejan Markovic, Simone Gambini, Rikky Muller, Michael Mark, David Chen, Will Biederman, Dan Yeager, Peter Ledochowitsch, Toni Bjorninen, Wen Li, Ping_chen Huang and Tsung-Te Liu to this presentation are gratefully acknowledged.
Research performed as part of the Berkeley-UCSF Center for Neural Engineering and Prosthetics. The support of the California Discovery program, the FCRP MuSyC and SONIC centers, and the member companies of BWRC is greatly appreciated.