Transcript (pptx)

Pulsed Neural Networks
Neil E. Cotter
ECE Department
University of Utah
Overview
Challenging Problems
Artificial Neural Networks
Learning
Dynamic Networks
Pulsed Networks
Applications
Challenging Problems
Challenging Problems
Image recognition in varied settings
Speech recognition in noise
Robotic control and navigation
Artificial Neural Networks
Biological Neuron
Biological Neuron
t0
w0
t
t1
S T
w1
t
t2
w2
t
t
tn
Perceptron
w0
x0
w1
x1
S
w2
x2
S
y
y
Perceptron Response
y(x1, x2)
x2
y=1
y=0
x1
Linear Separability
0 0
0
1
0
0
1
1
1
0
11
1
1
1
0
0
0
0
0
1
x2
1
Linear Separability
0 0
0
1
0
0
1
1
1
0
11
1
1
1
0
0
0
0
0
1
x2
1
Logic Gates
x1
1
0
(0, 1)
(1, 1)
y=1
y=0
0
0
(0, 0)
(1, 0)
x2
Parallel Processing
Sigmoid Neuron
w0
x0
w1
x1
S
w2
x2
S
y
y
Sigmoid Neuron Response
y(x1, x2)
x2
x1
Neural Network
u1
uL
Neuron
Neuron
wt S
Neuron
Neuron
wt S
Neuron
Neuron
wt S
Neuron
Neuron
wt S
y1
yM
Neural Network
u1
uL
Line
AND gate
wt S
Line
AND gate
wt S
Line
AND gate
wt S
Line
AND gate
wt S
y1
yM
Universal Approximation
y
x2
x1
Universal Approximation
y
x2
01
0
1
1
0
x1
Universal Approximation
y
x2
x1
Function Approximations
Learning
Gradient Descent
E(w1, w2)
w2
w1
Local Minima
E(w1, w2)
w2
w1
Backward-Error Propagation
E(w1, w2)
w2
w1
Dynamic Networks
Dynamic Networks
Neural Network
∫ dt
∫ dt
x1
xN
u1
y1
uL
yM
Learning through Time
State
0 or 1
1
v1
x
noise
v16
Adaptive Critic
p
reward = ±1
filter
adaptation
w1
x16
x xq q
noise
Associative Search
q
y
w16
±F
x
Learning through Time
State
0 or 1
x1
v1
noise
v16
x x qq
x16
w16
reward =
±1
l
decay
w1
Adaptive Critic
p
delay
a
b
noise
^r
Associative Search
q
y
±
F
Pulsed Networks
Pulsed Neuron
t0
t
w0
t
t1
t
S T
w1
t
t
t2
w2
t
t
t
tn
Pattern Processing
T
t
t0
t1
t2
Response Regions
0
x2
Perceptron
Pulsed Neuron
x3
x3
0
0
1
1
0
0
x1
x2
0
x1
Pattern Windows
t0
t1
t2
t
Pattern Windows
t0
t1
t2
t
Dynamic Pulsed Networks
Pulsed Neural Network
x1
xN
u1
y1
uL
yM
Applications
Speech Recognition
t
w
Sound Window
Spectrum
Auditory System
Basilar Membrane
Sound
t
t
Research Opportunities
Mathematically describe pattern processing
Devise pattern-learning algorithms
Build circuits
Title