& (3 ($ % $($5 # & / $!$ #
- )$#' # $" . "%! '
# 4 /
# , &' (/ $ ! $&# 6 & ! /
#$( & ##+ ! )# 2 / : 978:
#&'2$ % &+* * ) ( ) ')% * ( )+ ' + )# $ 3: &+ ) ') 0) $ & !& )!& & )'*+ * * ; < * 2 $$ * + + ) + &+ )9 *0((')+ ' + $! ')&! !* '1 )3 ()' ) %7 7 & + % % ) '%( &! * ' & 2 )% !* ) +$3 (() ! + 9
'&+)! 0+')*8 $ $'&7 '* )% & 7 2 ) & 7 ' &! +7
3*+ %* '& &'* $
& ')% ,'& )! *
! $ ) !47 9 $ & 7 9 & 0$37 & ##'& &7 " & )#'1! 7 !%'& % !&!7 !##3 0$$ )7 ! $ )#7 1! &7 !$$ ! )% &7 & )7 + ) ' '2!+* 7 '&! "')&!& &7 & !7 *0& : !07 9 #*!%'1! 7 9 )+) & 7 )0&' $* 0* &7 &. & )1 7 !& : & 0 & 7 * # & ')& 7 !$!( '& 7 0 *!* !+) 7 ** & $*7 1 & )% 7 ) * &
1
Energy (norm.)
12x
0.1
0.01
0.3V
0.001 0
0.2 0.4 0.6 0.8 VDD (V)
# $
1
1.2
0 )- # ! ) ' , ) 2 3 )! )!" ! 1
" )! " ) (
& " " ! $
!" & "&
4 / ). 977;5
!
4 / . . 978:5
# % # !
& " ()'
$
# ! " ) # + ,
- #' ( & .
! #* !
â&#x20AC;¯ / â&#x20AC;¯ # ' & #* ! & #&
' &
#&
# ) ! " 1 # # ! 2
Antenna Load (LED)
Impedance Matcher Voltage Doubler
3 &!# "). / ! "4
0 ! ) !' "# ! 3 / !"- 4
! ! ! ( #! "
â&#x20AC;Ż & #&! ) " ! $ â&#x20AC;Ż !) ! ! & $ ". " " - ! )0 !' "$ "# ! - !)- ! "" & $
". - ! #)-
3 &!# "). / ! "- / ! *4
E W T A H W T I S I N V O E I R T A O T T U P M O NEED C Y B N A E M
2-3 orders more efficient than todayâ&#x20AC;&#x2122;s silicon equivalent (>1016 FLOPS with ~20 W) Robustness in presence of component failure and variations â&#x20AC;Ż Neural response is highly variable (Ď&#x192;/Îźâ&#x2030;&#x2C6;1) [Faisal]
Amazing performance with mediocre components â&#x20AC;Ż E.g. sensory pathwaysâ&#x20AC;&#x201C; auditory, olfactory, vision,
!
& ' "
$ # # # ( $ # % ,*+-)
! # # % ! " #& # # " "&"# " $ " ' ! #! " ## ! "$
"#
Physical Interface Platforms across Scale and Modality
#! " ! # 1 ! . # & ! & #
. ! *
( #! # "
1 ! & # !) !# ( # ! ' #" . . ! *
# ! ' #" . & !
. ! *
ÎźECoG+BMI Peter Ledochowitsch / Aaron Koralek
Carmena / Maharbiz
" !$ ! $ " ! & #! / # #! " "
" #) ( #! "- #! 0 $
" . ! *
)" * !
. ! *
Free-floating wireless AP acquisition electrodes [Biederman, Yeager, VLSI12]
Spike Detection
Asynchronous 250 nW/ channel spike-sorting [Liu, VLSI12]
Preamble Buffer Feature Extraction
Spike Alignment
Register Bank Memory
"!#
64 channel remotely powered wireless uECoG [Muller, Le, Ledoschowicz, Li]
! %#$&"
! %#$&"
! % &(' μ
[DJ Seo et al, Arxiv, June 2013]
â&#x20AC;¯ $ â&#x20AC;¯
$ % & Voltage (mV)
20
Input Input Reconstructed
0
-20
-40
200
400
600
800
1000 1200 1400 1600 1800 2000 i d
20
0
-20
Slow conducting peaks
-40 50
' # )*(
100
150
200
250
300
#
& ! & ! ' ' ( ' ) ( ! )
! # )
( ")
$
!
! % !
! % ! ) ! â&#x20AC;¢â&#x20AC;¯ $ ! % â&#x20AC;&#x201C;â&#x20AC;¯ " % ! % â&#x20AC;&#x201C;â&#x20AC;¯ % " ( ) (
" â&#x20AC;&#x201C;â&#x20AC;¯ ( # ! ( ) "
â&#x20AC;¢â&#x20AC;¯ # ! ) % â&#x20AC;&#x201C;â&#x20AC;¯ ! ! " % ! % *, + â&#x20AC;&#x201C;â&#x20AC;¯ # % ! â&#x20AC;&#x201C;â&#x20AC;¯ ! " %
" % !
"
"
"
"
Functional non-determinism present in most applications related to human-cyber interfaces feature extraction, classification, synthesis, recognition, decision making, learning,
$ " $ ,
02/
! ( 0 0/+0// *
! # 0 1
# $ - $( ) ' .
"
!
!
Sparse Representation
Storage Recognition
Energy-efficient Analog Adaptive Front End
CNFET+RRAM Sparse Vector Generator
Correlation Matrix Memory
" !
!
Choose the right information representation that â&#x20AC;Ż makes computation easy, efficient and robust â&#x20AC;Ż matches the platform!
Recognition
Feature Extraction
Sensor Responses
$ & %
â&#x20AC;¢â&#x20AC;¯
& â&#x20AC;¢â&#x20AC;¯
"
! #
! Trainer Training Data
Model
â&#x20AC;&#x2DC;WEAKâ&#x20AC;&#x2122; Linear Classifier 1 Weight 1 Sensor Data
Weighted Voter
â&#x20AC;&#x2DC;WEAKâ&#x20AC;&#x2122; Linear Classifier 2 Weight 2
SENSOR ARRAY Projector
Weak TFT classifier branches on glass
MB-
1cm
8x8cm photoconductor plane on glass
0.8
tn
0.7
tp
0.6 0.5 0.4
(measured)
1
2
3
4
5
6
No. of weak classifiers
MSMB+ GND
0.9
0.3
THIN-FILM CLASSIFIER
MS+
Rates (tp & tn)
Weak classifier fault due to circuit non-idealities
SVM tp tn
1 â&#x20AC;&#x2DC;STRONGâ&#x20AC;&#x2122; Classification Result
DAQ SYSTEM / PC
2-5 weak classifiers based on highly non-ideal (TFT) multipliers achieve performance of ideal SVM
Probe card for testing
" #
" (
#
* +10 $ ,
j i
K ij
â&#x20AC;¯ #
â&#x20AC;¯ ! * M
â&#x2C6;&#x201A;θi â&#x2C6;&#x201A;t
= Ï&#x2030; i + â&#x2C6;&#x2018; K ij sin(θ j â&#x2C6;&#x2019; θ i ) j=1
- #( ) ' .
" # * , $ + â&#x20AC;¢â&#x20AC;¯ ! " ( ! $ $ â&#x20AC;¢â&#x20AC;¯
% - )
â&#x20AC;¯ â&#x20AC;¯ â&#x20AC;¯
& + ' 0 1 ! ! % - # . 09 54*4441 $
â&#x20AC;¯ & ' # â&#x20AC;¯ # ' ! ! * $ â&#x20AC;¯ # #
80
70
50
40
30
20
10
0 0
854444 85444 8544 854
60
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
$ /
â&#x20AC;¯ % # ! & ' $ # 6
2 + , $ * # ! #! # ! % / * 64473
1
' 1 "& ( ' , " ( % 76 % 6555 ! 1 % $ (
/ -0
%! !
'
' 65+555 ! & !
0
"& (
"& (
76+555 65+555 & ! ! "& (
" % ! % :8-9= "
2 % ! (, - & + 3
% ( ! + % + .
Learning
(Correlation)
Retrieval (Projection)
(Thresholding)
# ! ! ( & " '
â&#x20AC;¯ $ ) â&#x20AC;¯ & , # ! ) â&#x20AC;¯ - . *
# %
# # - # .
% *
37+54
/ 42360
# $
% CNFET RRAM
* +
# * +
22 100 Îźm Delay Cell
CNFETs
RRAM /
ILV
Target / > 1 5
0
Measured CNFET Ď&#x192;/Îź = 0.57 6
8
10 10 RRAM ( )
Delay Cell Coincidence Detector
RRAM ILVs
$ &
, $ ) .0-
100 Îźm
â&#x20AC;Ż / ! 1 & â&#x20AC;Ż 1 &
! $ !
+ ' * !
$
3.5
3
2.5
2
1.5
1
0.5
0 0
0.05
0.1
0.15
0.2
0.25
0.3
0.35
0.4
% &
ZN=0
ZN=1
Distribution
Up3Dn1
w/o IOFF with I OFF w/o IOFF with w/o
IOFF IOFF IOFF
with
Up3Dn2
Up1Dn1
VX (V)
% " ! " " !
! &
! &
% " $ '
'
! ' " ! #& " ! ! *()+"
, ) â&#x20AC;¯ , , $ # ! -â&#x20AC;¯ ! â&#x20AC;¯
â&#x20AC;¯
â&#x20AC;¯
* " + * " + -â&#x20AC;¯ $ $ " ' " $ ( # " * , + " $
!
! U O Y K N A H T MER
CI B
EAU C
OUP !