Future of neuromorphic systems Implantable medical electronics ...

bouncerarcheryAI and Robotics

Nov 14, 2013 (4 years ago)

78 views

1

Neuromorphic Engineering



University of Oxford

Department of Engineering Science

Natasha Chia


2

Outline


Definition of neuromorphic systems


Principles of neuromorphic technology


Typical applications of neuromorphic
technology


The basic computational element


Neural codes and information
representation


Future of Neuromorphic systems



Neuromorphic Engineering
3

Definitions


Carver Mead introducec the term
Neuromorphic Engineering

to describe

A new field of engineering whose design principles and architecture are biologically

Inspired.




Neuro

“ to do with neurons i.e. neurally inspired”



Morphic

“ structure or form”


Emulates the functional structure of neurobiological systems.



4

Principles of neuromorphic
technology

Build machines that have similar perception capabilities as human perception

Adaptable and self organising

Robust to changing environments




Realisation of future “THINKING” machines

(intelligent and interactive systems)


5

What does neuromorphic
Engineering involve?

Neuromorphic
Modeling
Neuromorphic
Computation
Analyze neurophysiological
functions in order to reproduce
neuronal structures and
architectures
Build neural networks
and rely on modellers for
explanation and modelling
of natural phenomena
6

Applications of
Neuromorphic Systems



N
EUROMORPHIC
ENGINEERING
E
LECTRICAL
E
NGINEERING
C
OMPUTER
SCIENCE
N
EUROSCIENCE
- Sensory systems
- Biorobots
- Neuron modelling
- Unsupervised learning
- Pattern understanding
7

Neuromorphic systems

Silicon Retina

Silicon Cochlea

TouchPad

Learning and adaptation

silicon systems

Koala
-
obstacle/tracking robot


8

Example of Modelling approaches


Parallel distributed analog correlation based processing
is the basis of VLSI systems that emulate the function
of neural information processing in biological systems

Silicon Cochlear Modelling
Biological/Physical
models
Analog VLSI
models
Digital VLSI
models
FPGA digital filters
IIR/FIR
Neuromorphic
models
Parallel processing circuits
Log domain ccts
Switched capacitor ccts
OTA biquads ccts
WLR ccts
1D BM model
2D hydrodynamic model
3D immersed boundary
model
9

The Basic computational element :
The Neuron

10

The Neuron Model

X(nT)
X(nT)
X(nT)
W_linking
W_feeding
W_inhibitary
T
T
T
Dynamic thresholding
Y(nT)










t
E
E
t
G
t
V
t
E
dt
dE
K
K
E





.

11

History of neuron modelling


1943


McCulloch & Pits


1963


Hodgkin & Huxley modelled axon of




giant squid.


1970


Kiang & Gerstein numerical analysis of




interactions in nerve

cells


1983


Cohan & Mpitson Deterministic chaos used



to describe behaviour of single neurons


1985


Meyer et al Discovery of electrical





interaction between neurons


1992


Bower Quantitative evaluation of




functional data


1993


Knopf and Gupta Fundamental





neural processing element


1995


Bressler Parallel processing of





information

12

Neural code and information
representation

How does spikes represent sensory information?




Models of neuron spiking mechanism


Stimulus Response features


Group behaviour




13

Stimulus Response features


Average firing rate


Position of each


neuronal discharge


Instantaneous firing


probability.
(A ganglion’s firing
rate depends on stimulation)


14

Spiking Neuron models


Integrate and fire

Neurons spike regularly in response

to an external current. Rate of

spiking increases with the magnitude

of the stimulus current


Stimulus Response
Model (SRM)















'
'
'
2
2
.
.
.
_
2
2
2
2
dxdt
t
t
A
x
B
t
x
I
N
t
Rate
Firing
e
k
e
k
x
B
he
t
t
A
s
c
r
x
s
r
x
c
t











Firing_R(t)
Time
A(t)
Space
B(x)
15

Neural coding


Information is conveyed to the brain in parallel by
spike trains of the nerve fibers.

Temporal pattern
coding
Channel
coding
Time of arrival
coding
Joint response
from multiple neurons
coding
t1 t2
Narrow correlations
Medium correlations
Broad correlations
16

Group behaviour of spikes

O
ne of the challenges has been to

understand the relative
contribution of various groups of spikes



O
scillatory coupling effects and amplitude dynamics in two or
more populations of neurons will be important topics for
future research


Some other examples: STDP(Spike Time
-
dependent plasticity)

Group neurons with correlated inputs


17

Future of neuromorphic systems

Implantable medical electronics

Increased human computer interaction

Intelligent transportantion systems


Learning, pattern recognition

Robot control(self motion estimation)

Learning higher order perceptual computation


18

Conclusion

Neuroscientists
and Biologists
Analyze neuronal
information
Electronic
Engineers
Reproduce
neurophysiological
phenomena in silicon
Computer Scientist
and Mathematicians
Modelling of
neuromorphic
systems