Cam brain

BUILDING AN ARTIFICIAL BUILDING AN ARTIFICIAL BRAINBRAIN

Using an FPGA CAM-Brain MachineUsing an FPGA CAM-Brain Machine

Mika ShoshaniMika Shoshani

Yossy SalpeterYossy Salpeter

An ARTIFICIAL BRAIN?!

• What?– A machine modeling the Human brain

• Why?– Breaking the limits of traditional computers

• And How?– “Teaching” the machine…

Scope

• Introduction

• Background– The basis of the “Brain Building” field

• The CAM-Brain machine

• Domo Arigato Mr. ROBOKONEKOROBOKONEKO– “Proof of concept”

• What’s Next...

Buzz words

• Neurons, Axons, Dendrites…

• Neural Network Module

• CAM - Cellular Automata Model

• FPGA - Field Programmable Gate Array

• Genetic Algorithms

• “Evolvable Hardware”

• A network of 1014 neurons• Data transfer by electric signals• DendriteDendrite cells (neurons Input)

– Collect signals and pass them to the neuron

• NeuronsNeurons– “Decide” when to initiate a signal

• AxonAxon cells (neurons Output)– Propagate neuron signals

The Human Brain

Genetic Algorithms

• A process imitating natural evolution

Random population

Fitness function

The fittestCrossover & MutationNew Generation

REPRODUCTION

Genetic Algorithms

• A process imitating natural evolution

Random population

Fitness function

The fittestCrossover & Mutation3’ed Generation

REPRODUCTION

Genetic Algorithms

• A process imitating natural evolution

Random population

Fitness function

The fittestCrossover & Mutation4’th Generation

REPRODUCTION

Genetic Algorithms

• A process imitating natural evolution

Random population

Fitness function

The fittestCrossover & Mutation5’th Generation

Fittest individual

REPRODUCTION

“Evolvable Hardware”• The Application of a Genetic Algorithm

on programmable hardware:Chip with

random circuitsMeasuring circuit

BestPerforming

circuits

Random MutationsNew Generationof mutant circuits

REPRODUCTION

Functioningcircuit

Evolve Hardware to perform a desired functionEvolve Hardware to perform a desired function

AT HARDWARE SPEEDS!!!

Human Brain vs. The Computer

• 1014 Neurons• Parallel Computing• Speed: 100+ M./sec.• Natural Evolution

• CPU - Central Processing Unit

• Serial Computing• Approx. Speed of light• “Designable”

The CAM-Brain Machine (CBM)

• A research tool of an artificial brain

• Consists of 32,768 neural modules

• Neural modules evolve in hardware using Genetic Algorithms

CBM Goal

• Create a complex functionality without any a priori knowledge of howhow to achieve it…

• Requires the desired Input/Output function!

CELLULAR automata MODEL

• A 3D grid of cells

• Each can be in one of a finite number of possible states.

• Sync. updated in discrete time steps.

• According to a local, identical interaction rule. “Chromosome”

CBM Neural Network Model

• The CBM implements the:

“CoDi” Cellular Automata based

neural network model

• Goals:– Fast evolution– Portability into electronic hardware

CoDI Cell design

• A cube with six neighbor cells

• Can function as Neuron, Axon or Dendrite

• A Neuron Cell:– 5 dendritic inputs + 1 axonic output– 4-bit input accumulator, “fires” on threshold

• A Dendrite cell: 5 Inputs / 1 Output

• An Axon cell: 1 Input / 5 Outputs

CoDI Module Evolving

• All cells are seeded with “chromosome”

• Seed Neuron cells randomly

• Growth procedure:– Each Neuron sends grow dendrite/axon signals– Blank cells become dendrite/axon– Grown cells propagate growth signals– Propagation direction is set by the chromosome

CoDI Module Evolving

CoDI Module evolution

• Each module is given a specific function

• Genetic Algorithem:– Initial population of 30-100 modules– Run for 200-600 Generations– Up to 60,000 different module evaluations

• Full module evolution takes approx. 1sec

CBM Architecture

• Cellular Automata Module

• Genotype/Phenotype Memory

• Fitness Evaluation Unit

• Genetic Algorithm Unit

• Module Interconnection Memory

• External Interface

Architecture {1}

• Cellular Automata Module– The hardware core of the CBM

– 3D array of identical logic circuits (cells)

– Module size of 24*24*24 cells (13,824)

– Implemented by 72 FGPAs

– Time shared between multiple modules - Forming a brain during simulation.

– No idle time between modules

Architecture {2}

• Genotype & Phenotype Memory– Total 1180 Mbytes RAM– Genotype memory for Evolution mode:

• Store Chromosome bitstrings• Store module neuron location & orientation

– Phenotype memory for Run mode:• Holds all evolved module maps

– Can support up to 32,758 modules

Architecture {3}

• Fitness evaluation unit – Evaluates module fitness– Signals each module inputs– Compares Module output to target output– This comparison gives a measure of

module performance

Architecture {4}

• Genetic Algorithm Unit – Selects a subset of the “best” evolved

modules for reproduction– Implements Crossover and Mutation masks– Generates offspring modules– Offspring chromosome generated in

hardware

Architecture {5}

• Module Interconnection Memory – Supports operation of Evolved modules as

one artificial brain– Provides signaling between modules

Architecture {6}

• External Interface – CBM Signaling is by 1-bit spiketrains– I/O For each module

• Input of up to 188 spiketrains• Output of up to 3 spiketrains

Human Brain vs. CAM-Brain

• 1014 Neurons• Parallel Computing• Speed: 100+ M./sec.• Natural Evolution

• 4*107 Neurons• 1150 parallel neurons• Approx. speed of light• “Designable” Evolution

• Political & Strategic goals

• A controlled cat as a “proof of concept”

• Radio connected to CBM

• Demonstrates CBM via evolved behaviors

• GoalGoal - The “CUTE” factor...

ROBOKONEKO

Behavior Evolving

• Moition control modules– Fitness criterion - speed & distance – Mechanical vs. Simulated behavior evolving– Slow evolution, 2-3 min. per chromosome– Hand coded base criterion.

• Non motion control modules evolution -Predicted to be Faster

SUMMARY

• Artificial Brain Building

• “CAM Brain Project”– Aims to build an artificial brain with 32000

evolved net modules, 40 million neurons

• “Robokoneko”– A Cat robot controled by the CAM-Brain– In development of motion control modules

What’s Next...

• “Intelligent” robotic pets, Household robots, Soldier robots.

• Artilect - Artificial Intellect

• Ultra-Intelligent Artilect = Moral dilemma

The prophecy

• Future WAR “Cosmists” vs. “Terrans”…

• The End of Human race as we know it...

References {1}

• "Building an Artificial Brain Using an FPGA Based CAM-Brain Machine", Applied Mathematics and Computation Journal, Special Issue on "Artificial Life and Robotics, Artificial Brain, Brain Computing and Brainware", North Holland. (Invited by Editor, to appear 1999), Hugo de Garis, Michael Korkin, Felix Gers, Eiji Nawa, Michael Hough.

• "A 40 Million Neuron Artificial Brain for an Adaptive Robot Kitten "Robokoneko", Hugo de Garis, Michael Korkin, Gary Fehr, Nikolai Petroff, Eiji Nawa, to be submitted to the Connection Science Journal, Special Issue on Adaptive Robots.

• "Simulation and Evolution of the Motions of a Life Sized Kitten Robot "Robokoneko" as Controlled by a 32000 Neural Net Module Artificial Brain", Hugo de Garis, Nikolai Petroff, Michael Korkin, Gary Fehr, Eiji Nawa, (Invitation by Editor to the Computational Geometry Journal (CGJ), Special Issue on Computational Geometry in Virtual Reality)

References {www}

• A Brief Introduction to Genetic Algorithms, by Moshe Sipper, http://lslsun.epfl.ch/~moshes/ga_main.html

• Non-uniform cellular automata, by Moshe Sipper, http://lslsun.epfl.ch/~moshes/ga_main.html

• Prof. Dr. Hugo de Garis Home Page, http://www.cs.usu.edu/~degaris/

• CNN - Swiss scientists warn of robot Armageddon, http://www.cnn.com/TECH/science/9802/18/swiss.robot/

• ,האוניברסיטה העברית בירושלים - המוחhttp://gifted.snunit.k12.il/activities/brain/