Erlang: The Unintentional Neural

Network Programming Language

Erlang Factory
San Francisco March 20ll
Gene Sher

lntroduction
!
The Long Standing Goals of Computational
lntelligence
!
Brain is just an organic substrate based NN
!
Blue Brain Project - http://bluebrain.epfl.ch/
!
Computer hardware is advancing steadily
!
A programming language is needed with the
right features
!
Erlang: As the NN Programming Language

Computational lntelligence
Through Genetic Algorithms and
Neural Networks

Biological Neural Network

Artificial Neural Networks
!
Simulate biological
NNs to various
degrees of precision
!
Directed graphs
!
Parallel
!
Learn, adapt, and
generalize
n-1 n-2
n-3
n-4 n-3
Sense/lnput
Act/Output

An Artificial Neuron
AF:tanh
Weights:
[1,0,-1|
Bias
[-0.76]
[0.46]
Output
lnput
lnput
[l]
l. Dot product:
DP=(l0.76) + (00.46)
Threshold = (-ll)
2. Activation strength:
Output = tanh(DP+Threshold)
[-0.94]

The lnput is Just a Vector
AF:tanh
Weights:
[0.3,0.2|
[-l,l]
Output
lnput
l. Dot product:
DP=(0.5-l) + (0.2l)
Threshold = (0l)
2. Activation strength:
Output = tanh(DP+Threshold)
[-0.29]

A Neural Network
[0.3,1| [1|
[1,0,-1|
[0.5]
[2,-2]
tanh(0.52 + l-2)
tanh(l0.5)
l
Bias
tanh(l-0.76 + 00.46 + -ll)
[-0.76] [0.46]
[-0.94]
Wind Speed Now
Wind Speed Later
Coordinates
tanh(tanh(C dot W) + tanh(WSP dot W))

NN Learning & Plasticity
Algorithms
!
Supervised
!
Backpropagation
!
...
!
Unsupervised
!
Kohonan (Self-organizing) map
!
Adaptive Resonance Theory
!
Hebbian
" "The general idea is an old one, that any two cells or systems of cells that are repeatedly active at the same
time will tend to become 'associated', so that activity in one facilitates activity in the other." (Hebb l949, p. 70)
!
Modulated
!
Evolutionary
!
..

Evolutionary Computation
!
Based on evolutionary principles
!
Stochastic search with a purpose
!
Survival of the fittest
!
Genotype to Phenotype
!
Mutation and crossover

George E. P. Box
!
lmproving productivity in a chemical process
plant at lmperial Chemical lndustries Ltd.
l.Vary a setting
2.See what happens
3.???
4.Profit!!!

Evolutionary Computation Flowchart
lnitialize the
population
Create offspring
through
random variation
Evaluate fitness
of each candidate
solution
Apply selection
algorithm
Terminate?
Yes
No

Simple Genetic Algorithm Example

Genotypes
A l00l
B 0000
C l0l0
D 0l0l
Phenotypes
Genotypes
lll0
0l00
l0l0
0l0l
Phenotypes
Genotypes
lll0
llll
l0l0
00l0
Simple Mutations
Genotypes
A l00l
B 0000
C l0l0
D 0l0l
Phenotypes
Genotypes
l0ll
ll0l
l0l0
0l0l
Phenotypes
Genotypes
l0ll
ll0l
l00l
llll
Crossover
Gen-l Gen-2 Gen-3
Gen-l Gen-2 Gen-3
Phenotypes
Phenotypes

Simple Hill Climber
AF:tanh
Weights:
[W|
l. [l]
2. [-l]
Output
lnput
tanh(lW)
lnitial W = l
l want: Output == 0
l. Outputl = tanh(ll) = 0.76
Output2 = tanh(l-l) = -0.76
2. Perturbation power!!!
Perturbation = -0.5
Try W = 0.5 = l - 0.5
Outputl = tanh(0.5l) = 0.46
Output2 = tanh(0.5-l) = -0.46
That's closer! New W = 0.5
3. Perturbation power!!!
Perturbation = +0.2
Try W = 0.7 = 0.5 + 0.2
Outputl = tanh(0.7l) = 0.60
Output2 = tanh(0.7-l) = -0.60
Not as good as before, New W = 0.5
4. Perturbation power!!!
Perturbation = -0.5
Try W = 0 = 0.5 - 0.5
Outputl = tanh(0l) = 0 !!!
Output2 = tanh(0-l) = 0 !!!
The right weight is 0.

Evolutionary Computation
Approaches
!
Genetic Algorithms (John Holland, 73-75)
!
Population of fixed length genotypes, bit strings, evolved through perturbation/crossing
!
Genetic Programming (John Koza, 92)
!
Variable sized chromosome based programs represented as treelike structures, with specially
crafted genetic operators
!
Evolutionary Strategies (lngo Rechenberg, 73)
!
Normal distribution based, adaptive perturbations (self-adaptation)
!
Evolutionary Programming (L. & D. Fogel, 63)
!
Like ES, but for evolution of state transition tables for finite-state machines (FSMs)

Topology and Weight Evolving
Artificial Neural Networks
!
Populations and Fitness Functions
!
Parametric mutation operators
!
Topological mutation operators
!
Other mutation operators
!
Learning Algorithms, Activation Functions
!
Submodules

TWEANN Cycle
Seed NN population
Apply to
problem
Calculate fitness
scores
Select fit
organisms
Create
offspring

NN Programming Language

Necessary Features for a NN-PL
(These will sound very familiar)
!
Encapsulation
!
Concurrency through Neuron primitives
!
Fault detection primitives
!
Location transparency
!
Dynamic code upgrade

Erlang's Features
!
Encapsulation primitives
!
Concurrency
!
Fault detection primitives
!
Location transparency
!
Dynamic code upgrade

Neural Networks Through Erlang
Neural Networks Through Erlang

The Topological l:l Mapping
PId n-2
n-3
PId n-3
PId
PId
PId
N1 n-2
n-3
N4 n-3
N2
N3
N3
PId PId
Sl S2
Pld Pld Al A2

Monolithic NN Supervision Tree
NN_Sup
Supervisor
Sensor Actuator Neuron Neuron Neuron
Population
Monitor
Supervisors
Workers
Monitor Root
Supervisor
NN

States & Functions
!
Neuron
!
{lnputPlds,OutputPlds,Ws,AF,PF}
!
{U_Ws,Output} = update_weights(lnput,Ws,AF,PF)
!
fanout(OutputPlds,Output)
!
Sensor
!
{Function,OutputPlds,Parameters}
!
Sensory_Signal = sensor:Function(Parameters)
!
fanout(NPlds,Sensory_Signal)
!
Actuator
!
{Function,lnputPlds,Parameters}
!
actuator:Function(Actuator_Signal,Parameters)

Genotype Storing and Encoding
!
Tuple encoded
!
{neuron,lnputPlds,OutputPlds,Ws,AF,PF}
!
Relational database friendly
!
Human readable
!
Standard directed graph vertex and edge based
!
Mutation operators are standard digraph
opreators
!
Add/remove Neuron/Vertex
!
Add/remove Synapse/Edge

Mnesia as Storage for Genotypes
!
Robust and safe
!
Tuple friendly
!
Easy atomic mutations
!
lf any part of the mutation fails, the whole
mutation is just retracted automatically

Modular NN Topology Mapping
PId PId
PId
PId PId
PId PId
Pld Pld
N1 N2
N3
N4 N3
I1 I2
Ol O2
PId
Ml
M3
M4
M1 M2
Al A2
S1 S2

Modular NN Supervision Tree
NN_Sup
Neuron Neuron
Neuron
Population
Monitor
Supervisors
Workers
Modular NN
Module
Neuron Neuron Neuron
Module Supervisors
Root Supervisor
Sensor Actuator
Workers

What Erlang Offers to the Field of
NN Research
!
Augmenting topologies live
!
Full distribution and utilization of hardware
!
Fault tolerance; "stroke recovery"
!
l:l mapping, from ideas to prototype systems
!
No need to overcome linguistic determinism
!
Switching and adding new modules, no matter what
they do, requires no rewrite thanks to message
passing
!
Flexibility... in everything!

DXNN: A Case Study

Memetic Algorithm Based
TWEANN
Seed NN population
Apply to
problem
Calculate fitness
scores
Select fit
organisms
Create
offspring
Local Search:
Hill Climber

Evolving Topologies
N(1,1)
N(2,1)
S1
Al
N(1,1)
N(2,1)
N(3,1)
S2
Al
N(1,1) N(1,2)
N(2,1)
N(3,1)
S1 S2
Al
N(1,1) N(1,2)
N(2,1)
N(3,1) N(3,1)
S1 S2
Al A2

Modular DXNN Architecture
Cortex
M3
M4
M1 M2
Sensors
Actuators
OS

Modular DXNN Supervision Tree
Cortex
Neuron Neuron
Neuron
Population
Monitor
Supervisors
Workers
DXNN
Module
Neuron Neuron Neuron
Module Supervisors
Root Supervisor
Sensor Actuator
Workers

Seed Population
!
Total first layer neurons = total # of sensors
!
Total last layer neurons = output vector length
!
This is usually the sum of actuator vector lengths
!
Choose AF, PF... randomly from the available
lists

Complexification and Elaboration
!
Start with a simple initial topology
!
Add to and elaborate on the topology during
mutation phases
!
Apply parametric mutations only to the newly
created Neurons
!
Scale the fitness scores based on NN size

DXNN Genotype
!
{neuron,ld,lnputPldPs,OutputPlds,AF,PF}
!
lnputPldPs: [{Pldl,Wl}...{Pldn,Wn}]
!
{module,ld,lnputPlds,FanOut,Fanln,OutputPlds}
!
{cortex,ld,Sensors,FanOut,Fanln,Actuators}
!
{dxnn,ld,Cortexld,Modulelds,Neuronlds}
!
{available_AFs,[AFl...AFn]}
!
{available_PFs,[PFl...PFn]}
!
{available_SensorTypes,[Sl...Sn]}
!
{available_ActuatorTypes,[Al...An]}
!
{available_MutationOperators,[MOl...MO2]}

DXNN mutation operators
!
Local search mutation operators
!
perturb_weights(Weights)
!
Global search mutation operators
!
add_subcore(Genotype)/remove_subcore(Genotype)
!
add_sclink(Genotype)/remove_sclink(Genotype)
!
add_neuron(Genotype)/remove_neuron(Genotype)
!
add_nlink(Genotype)/remove_nlink(Genotype)
!
add_bias(Genotype)/remove_bias(Genotype)
!
change_af(Genotype)/change_plasticity(Genotype)

DXNN Platform
Cortex
M3
M4
M1
M2
Sensors
Actuators
OS
Camera
Image/Sensor
Data
Pan/Tilt
Control
Signals

DXNN Benchmarks & Application
Areas
!
Double pole balancing
!
Simple food gathering
!
Dangerous food gathering
!
Predator vs prey simulations
!
Room navigation and new sensor acquisition
!
Circuit design
!
Time series analysis
!
...

Double Pole Balancing Setup
N1 N2
N3
N4 N3
Cortex
[Force|
Actuators
Cart
[PlAngle,P2Angle,CPos]
Sensors

Double Pole Benchmark
Double pole balancing with velocities Without velocities, without damping Without velocities, with damping
Method Evaluations Method Evaluations Method Evaluations
RWG 474329 RWG 4l5209 RWG l232296
EP 307200 EP EP
CNE 22l00 CNE 76906 CNE 87623
SANE l2600 SANE 262700 SANE 45l6l2
Q-MLP l0582 Q-MLP Q-MLP
NEAT 3600 NEAT NEAT 6929
ESP 3800 ESP 7374 ESP 26342
954 l249 34l6
CMA-EX 895 CMA-EX 352l CMA-EX 606l
DXNN 725 DXNN 2359 DXNN 23l3
CoSyNE CoSyNE CoSyNE
Non DXNN data taken from Table-3, Table-4 of: Faustino Gomez, Jurgen Schmidhuber, Risto Miikkulainen,: Accelerated Neural
Evolution through Cooperatively Coevolved Synapses. Journal of Machine Learning Research 9 (2008) 937-965

Artificial Life Setup
DXNN
DXNN
DXNN
DXNN
Collision
Detection
Population
Monitor

2d Robots and Sensor Types
0-
90
90-
l80
l80-
270
270-
360
Predator
P
r
e
y
Sensor
angle
coverage,
resolution=4
Plant
Poison
Ray casting
based sensors

Predator vs. Prey

Evolving Hardware
PId PId
PId
PId PId
not and
or
and not
PId PId Sl S2
Pld Pld Ol O2
AFs: [tanh,sin,abs...]
AFs: [AND,NOT,OR]

Modules as general programs
Where each Module can be externally evolved NN, a general
program, a circuit, a substrate encoded NN...
PId
PId
PId
PId
Al A2
Sl S2
M3
M4
M1
M2
Al A2
Sl S2

QuadCopter Stabilization
Cortex
M3
M4
M1
M2
Sensors
Actuators
OS
Sensor Data
Prop Speeds: [P1,P2,P3,P4|
Player/Gazebo

Bipedal Gait
Cortex
M3 M4
M1
M2
Sensors
Actuators
OS
Sensor Data
R Joint Angles: [R1,R2,R3|
Player/Gazebo
L Joint Angles: [L1,L2,L3|

Time Series Analysis
(Through a sliding window)
Cortex
M3
M4
M1
M2
Sensors
Actuators
OS
l 2 l 4 3 4 2 3 2 2 3...
3.9
Prediction
Cur. Window

Coevolution
!
Environment and fitness landscape is created by
the interaction of competing species
!
Arms race

UAVs & Aerial Combat
Bently Creative Evolutiony Systems, Ch-l9, Discovering Novel Fighter Combat Maneuvers: Simulating Test Pilot Creativity
Cortex
Sensors
Actuators
Sensor Data
Pspeeds, Atk: [P1,P2,P3,P4,A|
Player/Gazebo
Cortex
Sensors
Actuators
Pspeeds, Atk: [P1,P2,P3,P4,A|
Sensor Data

Cyberwarfare
(Something like metasploit, but NN provides the parameters and chooses the tools)
DETER (NetSecTestbed) can be found at http://www.isi.deterlab.net/
Cortex
Sensors
Actuators
Network Data
Parameters: [T1,P1,T2,P2...|
ns3 or DETER
Cortex
Sensors
Actuators
Parameters: [T1,P1,T2,P2...|
Network Data
Attacker
Defender

Conclusion & Future Work
!
Memetic vs Genetic
!
Spiking neural networks and fault tolerance
!
Over-the-net distribution and sensing
!
Artificial life, 3d systems, distributed
processing...
!
Towards more realistic simulators and
evolving greater complexity...

Thanks! Questions?
!
Learn More
!
Preprints available on arxiv.org
!
Get the code
!
Will be available on github soon
!
Get in touch
!
[email protected]

END