Artificial Neural Networks Examples - cvut.cz · 01001110 01100101 01110101 01110010 01101111...

01001110 0110010101110101 0111001001101111 0110111001101111 0111011001100001 0010000001110011 0110101101110101 0111000001101001 0110111001100001 0010000001101011 0110000101110100 0110010101100100 0111001001111001 0010000001110000 0110111101100011 0110100101110100 0110000101100011 0111010100101100 0010000001000110 0100010101001100 0010000001000011 0101011001010101 0101010000101100 0010000001010000 0111001001100001 0110100001100001 00000000

Artificial Neural Networks Examples

Jan Drchaldrchajan@fel.cvut.cz

Computational Intelligence GroupDepartment of Computer Science and Engineering

Faculty of Electrical EngineeringCzech Technical University in Prague

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Outline

● Learning artificial neural networks (ANNs).

● Task to solve with ANNs.

● ANN applications.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

ANN, Learning & Recall

● ANN is a black box performing transformation.

● ANNs work most frequently in two phases:– Learning phase – adaptation of ANN's

internal parameters.

– Evaluation phase (recall) – use what was learned.

•••

•••

Artificial

Neural

Network

Input vectors Output

x1

x2x3

xn

y1y2

ym

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Supervised learning

● Learning by examples:

– given a set of example pairs Pi=(x

i,y

i),

– find transformation f which approximates y

i=f(x

i) for all i.

http://sugiyama-www.cs.titech.ac.jp/~sugi/figs/supervised-learning.png

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Unsupervised learning

● Self-organization, no teacher.● SOM, ART...

http://en.wikipedia.org/wiki/Self-organizing_map

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Reinforcement learning

● Teaching examples not available → they are generated by interactions with the environment (mostly control tasks).

http://anji.sourceforge.net/polebalance.htm

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Reinforcement Learning Demonstration

● 1994, Karl Sim's: Evolving Virtual Creatures video.

● Evolves both creature bodies (including sensors) and controlling networks...

● You can play with http://www.framsticks.com/

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Financial Applications

● Stock market time series forecasting.● Buy/sell timing detection and stock portfolio

selection.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Medical

● EEG/ECG processing – e.g. sleep disorder● Survival analysis – e.g. breast cancer● Pattern/Image recognition – MR, CT

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Compression by ANNs

http://www.cse.unr.edu/~looney/cs773b/NNimage-compress.pdf

cut here

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Neocognitron

Prof. Fukushima (1980)handwritten character recognition

similar to structures in brain

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Convolutional Neural Networks: LeNet

● Yann LeCun

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

CNNs: MNIST

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

CNNs: ImageNet

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

CNN for Speech Recognition

● Abdel-Hamid et al. (2014)

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

LSTM for Translation

● Sutskever et al. (2014)

● Long-Short Term Memory

● English → internal representation → French

http://arxiv.org/abs/1409.3215

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

CNN & LSTM for Activity Recognition

● Baccouche et al. (2011).

Baccouche, Moez, et al. "Sequential deep learning for human action recognition." Human Behavior Understanding. Springer Berlin Heidelberg, 2011. 29-39.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Playing Atari 2600 Games

● Mnih et al. (2015).

Mnih, Volodymyr, et al. "Human-level control through deep reinforcement learning." Nature 518.7540 (2015): 529-533.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Playing Atari 2600 Games

● Mnih et al. (2015).

Mnih, Volodymyr, et al. "Human-level control through deep reinforcement learning." Nature 518.7540 (2015): 529-533.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Playing Atari 2600 Games

● Mnih et al. (2015).

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Speechreading (Lipreading)

● Günter Mamier, Marco Sommerau & Michael Vogt, Universität Stuttgart.

● A neural classifier detects visibility of teeth edges and other attributes.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Detection and Tracking of Moving Targets (ICBMs)

● The moving target detection and track methods here are "track before detect" methods.

● They correlate sensor data versus time and location, based on the nature of actual tracks.

● The track statistics are "learned" based on artificial neural network (ANN) training with prior real or simulated data.

● Reduce false alarm rates by up to a factor of 1000 based on simulated SBIRS data for very weak ICBM targets against cloud and nuclear backgrounds.

http://tralvex.com/pub/nap/#Detection and Tracking of Moving Targetshttps://web.archive.org/web/20011204214308/http://www.ca.defgrp.com/detect.html

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

ANNs and Game AI

● AI opponents.● Modeling player behaviour.● Strategy estimation.● Realistic motion.● Cheat detection.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Fields of Battle (1995)

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Battlecruiser 3000AD (1996)

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Creatures (1996-)

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Blondie 24 Chess (1999)

● David B. Fogel.

● Combination of minmax. & neuroevolution.

● Coevolution.

● Defeated 99.61% of 165 online players (including masters).

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Colin McRae Rally 2.0 (2001)

http://www.ai-junkie.com/misc/hannan/hannan.html

● Opponents AI.● MLPs, RPROP learning.● ANN driving model follows

optimal track. ● Different models for different cars and

road conditions● Different networks for steering and

speed control.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Black & White (2001)

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Forza Motorsport 2 (2007)

● Opponent AI.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Parrot Speech (FEE CTU)

● Classification of parrot sounds → parrot speech consists of 41 “words”. Self Organization Map (SOM).

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Colabroute (FEE CTU)

● GPS Data Mining - THSOM.

● Automatic detection of crossroads.

● Detection of interesting places (gas stations, dangerous crossroads,...)

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

ShapeSOM (FEE CTU)

● Reconstruction of a room model out of unordered clusters of points.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Bone Age Modelling (FEE CTU)

● Modelling of age based on bone measurements.

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

Fetal Weight Prediction (FEE CTU)

● EFW = 0,0504AC2*16,427AC + 38,867FL + 284,074

A4M33BIA 2016

Jan Drchal, drchajan@fel.cvut.cz, http://cig.felk.cvut.cz

RoboNEAT (FEE CTU)

● Neuroevolution of robotic controllers.● HyperNEAT – large-scale ANNs.