(P-ITEEA-0033) Embedded Systemsusers.itk.ppke.hu/~fulta/d46/PITEEA0033_2016_Lecture_10.pdf · •...
Transcript of (P-ITEEA-0033) Embedded Systemsusers.itk.ppke.hu/~fulta/d46/PITEEA0033_2016_Lecture_10.pdf · •...
Pázmány Péter Katolikus Egyetem
Információs Technológiai Kar
Building blocks of bio-inspired
learning algorithms
Part 1/2
Lecture 10
May 4, 2016
Beágyazott elektronikus rendszerek
(P-ITEEA-0033)
Embedded Systems
Outline
• Example for complex space-time algorithm:
visual attentional selection, search and
classification
• Classification using neural networks
(solutions, feature extraction, etc.)
• Adaptive Resonance Theory (ART) network
(operation and application)
May 4, 2016 Lecture 10, P-ITEEA-0033
High level
autonomous
flight control
Controller
control
commands
flight
plan
SensorsSensory
information
processing
Example: unmanned aerial
navigation
May 4, 2016 Lecture 10, P-ITEEA-0033
Landscape patterns to recognize
May 4, 2016 Lecture 10, P-ITEEA-0033
Search and recognition
May 4, 2016 Lecture 10, P-ITEEA-0033
Tasks to solve
• Navigation (to complete mission)
• Reconnaissance (based on land features)
• Processing video stream
– sensing, extracting space-time features
• Recognition, decision
• Maintaining contact (base, other UAVs)
• Sensing, computing, communication ...
• Speed, power and physical limits!May 4, 2016 Lecture 10, P-ITEEA-0033
Sensing-processing for
exploration/selection/tracking/navigation
CMOS sensor(cut through an image
window controlled by an
attention mechanism)
CNN sensor-processor(exploration by parallel,
spatio-temporal nonlinear
feature extraction)
DSP (selection/ tracking/
navigation)
…
Sensor Image
1024x1024
Window
128x128
Nonlinear
spatio-temporal
channels
Global and
local feature
descriptors
Selection of focus and
scale of attention
Optical flow &
navigation parameter
estimation (YPR)
Feature classification
using ART network
Upper-level framework
(at different time-scales)
Y,
P, R
C
F, S
M
T
T
May 4, 2016 Lecture 10, P-ITEEA-0033
Integer DSP
(TX C6415
720 MHz)
Communication
Processor
(ETRAX 100)
Memory
(Flash &
SDRAM)
Color
CMOS Sensor
(IBIS 5-C)
Sensor-
processor
(ACE16k)
PLD (XILINX)
Ethernet /
RS 232 /
USB / Dig I/O /
FireWire
(1280x1024) (128x128)
Float DSP
(TX C6701
150 MHz)
Visual
input
Bi-i as embedded system
May 4, 2016 Lecture 10, P-ITEEA-0033
Pázmány Péter Katolikus Egyetem
Információs Technológiai Kar
Decision/classification with
neural networks
Classification methods using
neural networks
• Classifications using adaptive methods
• Input-output data representation
• Multi-layer perceptron (MLP)
• Supervised and unsupervised learning
• Off-line and on-line learning
• Categorization vs. classification
• ART and ARTMAP networks
May 4, 2016 Lecture 10, P-ITEEA-0033
Biological motivations
• Structure of typical nerve cell:
– body, dendrites, axon
• Interconnection
• Operation:
– excitation, inhibition, signal integration, firing
• Learning
May 4, 2016 Lecture 10, P-ITEEA-0033
The model neuron
…
f x( )
x1
x2
x3
xn
x0=1w1
w2
w3
wn
w0
y
May 4, 2016 Lecture 10, P-ITEEA-0033
The Perceptron
Function:
yx wi i
i
n
1 0
10
if
otherwise
where
and
x x i n
w R i ni
i0 1 1 1 1
0
, , , ,...,
, ,...,
‘net input’
May 4, 2016 Lecture 10, P-ITEEA-0033
Perceptron decision regions
x1
xn
May 4, 2016 Lecture 10, P-ITEEA-0033
Multi-layer Perceptrons (MLP)
• Architecture
• Combining decision regions
• MLP’s can implement arbitrary
decision/classification tasks
• Learning??
May 4, 2016 Lecture 10, P-ITEEA-0033
The ADALINE
• ADAptive LInear NEuron
• Output function is linear: f(x) = x
• Learning by minimising Mean-Squared
Error (MSE)
• The “delta” learning rule:
wE
wd y xi
ii
( )
May 4, 2016 Lecture 10, P-ITEEA-0033
The LMS training algorithm
1. Initialise weight vector (w = w0)
2. Select new training pair (x, d)
3. Calculate actual output y
4. Calculate error:
5. Adjust weight vector:
6. If MSE = min. Then stop Else Goto 2.
w d y f net x i ni i ( ) ( ) , ,...,0
d y
May 4, 2016 Lecture 10, P-ITEEA-0033
Minimizing errorError
Weight
May 4, 2016 Lecture 10, P-ITEEA-0033
Back-propagation networks
• Multi-layer net, nonlinear output function
• Solving the credit assignment problem
• Calculating error for hidden nodes (by
propagating error terms backwards)
• Minimising MSE by gradient descent
• BP nets are universal function
approximators
May 4, 2016 Lecture 10, P-ITEEA-0033
Back-propagation training
…
…
…
Error
InputOutput
TargetError
May 4, 2016 Lecture 10, P-ITEEA-0033
Problems with BP training
• Only local minima can be found
• Overtraining (overfitting)
• Long training time
• Catastrophic forgetting
May 4, 2016 Lecture 10, P-ITEEA-0033
Learning by self-organisation
• Learning without a teacher
• Typical network architectures and learning
methods
• What can be learned?
• Learning classification?
May 4, 2016 Lecture 10, P-ITEEA-0033
Self-organising networks
…
…
x1 x2 x3 xn
y1 y2 ym
May 4, 2016 Lecture 10, P-ITEEA-0033
Self-organizing networks
• Winner-take-all at output
• Unsupervised learning
yD w x D w x i j
ii j
1
0
if
otherwise
( , ) ( , ),
wx w i
iiold
( ) if node wins
otherwise0
May 4, 2016 Lecture 10, P-ITEEA-0033
Unsupervised learning example
May 4, 2016 Lecture 10, P-ITEEA-0033
Adaptive Resonance Theory
(ART) networks
• Stephen Grossberg (1976, 1987, …)
• Gail Carpenter (1987, …)
• Goal:
“Autonomous learning within complex
environments that are not under strict
external control.”
• The Stability-Plasticity dilemma
May 4, 2016 Lecture 10, P-ITEEA-0033
ART networks (cont’d)
• Human capabilities:
– ability to pay attention (at variable levels)
– expectation & reaction to unexpected events
– learning without a teacher
• ART is capable of …
“development of stable recognition codes
by self-organisation in real-time in
response to arbitrary sequences of input
patterns.”
May 4, 2016 Lecture 10, P-ITEEA-0033
ART variants
– ART1 (1987): binary input patterns
– ART2 (1987): continuous (real-valued) inputs
– ART3 (1990): hierarchical search
– ARTMAP (1991): supervised ART
– Fuzzy ART(MAP) (1991, 1992): ART and fuzzy
logic
– …
May 4, 2016 Lecture 10, P-ITEEA-0033
Definition of ART networks
• Level 1: Nonlinear differential equations
(activation, coupling, signal propagation &
learning laws)
• Level 2: Asymptotic approximation using
algebraic equations
• Level 3: Algorithmic description of
emerging behaviour as a result of network
interactions
May 4, 2016 Lecture 10, P-ITEEA-0033
ART network architecture
F1 STM
STMF2
Gain Control
Gain Control
Input pattern
LTM
LTM
Attentional Subsystem Orienting
Subsystem
A
STM
Reset
wave
May 4, 2016 Lecture 10, P-ITEEA-0033
ART network dynamics
...
F0
F1
F2
Attentional subsystem Orienting
subsystem
May 4, 2016 Lecture 10, P-ITEEA-0033
ART network dynamics
1 0 1 1 1 0
?
May 4, 2016 Lecture 10, P-ITEEA-0033
ART network dynamics
?
!
1 0 1 1 1 0May 4, 2016 Lecture 10, P-ITEEA-0033
ART network dynamics
May 4, 2016 Lecture 10, P-ITEEA-0033
ART network operation
Compute the similarity
with the Choice function
Select the best prototype
Winner-Take-All strategy
Check the difference
Match functionr
Update the winner
Learning function
Create
new category
I
Categories
pass
winner
Resetfail
May 4, 2016 Lecture 10, P-ITEEA-0033
ART1 example
InputCategory prototypes
1 2 3
5.0r1 2 3 4 5
8.0r
May 4, 2016 Lecture 10, P-ITEEA-0033
ART1 emergent behaviour
The ART1 network implements a fast and
stable incremental clustering algorithm.
May 4, 2016 Lecture 10, P-ITEEA-0033
Properties of ART• stability vs. plasticity:
– attentional system: processing of familiar inputs
– orienting subsystem: controlling search in case
of unfamiliar inputs
• prototypes with self-scaling critical features
• direct access to categories after self-
stabilisation
• attentional vigilance (can be modulated)
• learning in resonant state
May 4, 2016 Lecture 10, P-ITEEA-0033