Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of...

17
Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 mplementation of Artificia ntelligence System on FPGA
  • date post

    15-Jan-2016
  • Category

    Documents

  • view

    214
  • download

    0

Transcript of Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of...

Page 1: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Presenting: Itai Avron

Supervisor: Chen Koren

Final Presentation

Spring 2005

Implementation of ArtificialIntelligence System on FPGA

Page 2: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Project Goals

• Creating a VHDL design of a Neural Network

• Comparison Vs. software implementation (Matlab)

Page 3: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Background• Neural Network is a Learning Machine

• It is build from Neurons (Perceptrons), which holds the knowledge of the system within their inter-connection strength

• Every Neuron Implement the Active Function:

0

1

wxw i

n

i

Ti

Page 4: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

System Interface

• Input: - Image (16x16 pixels) - Weights

• Output: - A number between 0-9 (4 bit vector)

Page 5: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

System Architecture

Page 6: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Net Architecture

Page 7: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Controller – Flow Diagram

Page 8: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Neuron Architecture

Page 9: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Hardware Requirements

• Neuron : ROM – 2^15*20 bit = 80KB 257 Multipliers (20 bit input)

256 Adders (40-48 bit input)

• Network : Memory – Used : 17*(256+1)*20 bit = 10.7KBIn Reality : 32 Lines => 20.1KB

• System : Image registers – 20*256 bit = 640B

Page 10: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Simulations

Neuron:

Page 11: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Neural Network

Page 12: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Neural Network (Cont.)

Page 13: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

System

Page 14: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

System (Cont.)

Page 15: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Synthesis

• Synthesis on 3 different FPGA

1. xc2v1000-5-bg575 -> 67.128MHz

2. xc2v1000-6-fg256 -> 80.540MHz

3. xc2vp20-6-fg676 -> 75.603MHz

Frequency = 80MHz

Page 16: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Comparison

• Matlab : 114 errors out of 1000 pictures

Calculation time: 0.5970 sec

• VHDL : 114 errors out of 1000 pictures

Calculation time:

1000*43/80MHz = 0.5373 msec

The hardware is about 1000 times faster!!

Page 17: Presenting: Itai Avron Supervisor: Chen Koren Final Presentation Spring 2005 Implementation of Artificial Intelligence System on FPGA.

Improvement Suggestion

• Change numbers resolution to less than 18 bit(max input bits in Xilinx components)

• Implement learning is HW