alex acm sc machine learning day [materials] | machine learning& artificial intelligence ... our...

Upload: alex-acm-sc-library

Post on 14-Apr-2018

220 views

Category:

Documents


0 download

TRANSCRIPT

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    1/69

    Artificial Intelligence

    Our Ultimate Dream

    Mohammad Abdelaziz Gowayyed

    May 2013

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    2/69

    Agenda

    Introduction

    Example: How to build a machine to play

    chess?

    Artificial Intelligence Subfields

    Feature Extraction for Classification

    Artificial Neural Networks

    Future of Artificial Intelligence

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    3/69

    What is Artificial Intelligence?

    The study of computer systems that attempt

    to model and apply the intelligence of the

    human mind

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    4/69

    Artificial Intelligence Directions

    Build Systems that:

    act like humans

    act rationally

    think like humans

    think rationally

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    5/69

    Act like human: The Turing Test

    Turing Test:

    If a man had a conversation with a machine and

    he couldnt tell, whether it was a machine or a

    human, this machine passes the Turing Test

    It means that the machines perfectly acts like a

    human

    Does it need to think like a human to act like ahuman?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    6/69

    Think like human: Cognitive Science

    We must somehow determine first how

    human think, do we know?

    Introspection

    Psychological experiments

    Cognitive Science: construct precise and

    testable theories of the workings of humanminds through computer models and

    psychological experiments.

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    7/69

    Think rationally: Laws of Thought

    Formalizing logic

    Main emphasis is on inference

    Example: Socrates is a man; all men are

    mortal; therefore Socrates is mortal.

    By 1965, programs existed that could, givenenough time and memory, take a description

    of a problem in logical notation and find thesolution to the problem, if one exists.

    Problems?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    8/69

    Act rationally: The rational agent

    Rational behavior: doing the right thing

    The right thing: that which is expected to

    maximize goal achievement, given the

    available information

    Take decisions

    Reinforcement Learning: learn fromenvironment feedback based on my choices

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    9/69

    Is there a benchmark for humanintelligence?

    May be Chess?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    10/69

    How to program a machine to play

    Chess?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    11/69

    How to program a machine to play

    Tic Tac Toe?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    12/69

    Easy?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    13/69

    What is the difference betweenChess and Tic Tac Toe?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    14/69

    The general idea is to build a treerepresenting different possible paths,

    each path represents sequence of

    choices from you and your opponent.The tree will help the computer

    choose

    That is called aGame Tree

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    15/69

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    16/69

    Min-Max Algorithm

    Two agents are playing

    Zero-Sum Game

    Any loss to you is an equal win for your opponent

    Assume that you are playing with a rational

    player; he will try to maximize his profit

    At each step, you want to maximize your profitand assume that the opponent will minimize it

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    17/69

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    18/69

    Min-Max Algorithm

    Is it feasible to create the whole tree? Is it finite?

    What about the depth?

    Will we be forced to choose random moves?

    How to evaluate each state? Heuristic function (e.g. number of safe pieces)

    Different algorithms for traversing the tree Depth first

    Breadth First A*

    Do we need to visit the whole tree? Alpha-Beta pruning

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    19/69

    Heuristic Function

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    20/69

    Can we say that the computer is

    now intelligent because it candefeat human in chess?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    21/69

    There are still a lot of tasks that

    human does easily while it is very

    hard for machines

    Can you name a few?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    22/69

    Chinese Room

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    23/69

    Chinese Room

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    24/69

    Machine vs. Human

    Machine

    Easy Tasks

    Multiply 4878 * 1254

    Play chess Hard Tasks

    Understand concepts

    Recognize faces

    Recognize voices

    Human

    Easy Tasks

    Recognize faces

    Recognize voices Understand concepts

    Hard Tasks

    Multiply 4878 * 1254

    Play chess

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    25/69

    Some Artificial Intelligence Subfields

    Machine Learning

    Let machines learn rules (e.g. to classify or tochoose between alternatives Reinforcement

    Learning) Computer Vision

    Teach machines how to recognize visual world

    Natural Language Processing

    Teach machines how to recognize humanlanguages

    Speech Recognition

    Teach machines how to recognize human voices

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    26/69

    Some Computer Vision Tasks

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    27/69

    Face Detection

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    28/69

    Image Classification

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    29/69

    Human Action Recognition

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    30/69

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    31/69

    Feature Representation

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    32/69

    Feature Representation

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    33/69

    How is computer perception done?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    34/69

    Feature Representation

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    35/69

    Computer Vision Features

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    36/69

    Audio Features

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    37/69

    NLP Features

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    38/69

    Sensor Representation in Human Brain

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    39/69

    Can we mimic what already happens in

    human brains?

    The human brain consists of a network of

    neurons.

    Human brains learns to recognize images by

    the same method as it learns to recognize

    voices.

    Scientists still dont know how the brain

    works.

    Scientists only know some concepts that brain

    applies (e.g. sparse and general

    representations)

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    40/69

    Human Nervous System

    Around 85,000,000,000neurons each one

    connected to around 1000 others

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    41/69

    Number of neurons in elephant brains

    11,000,000,000 neurons

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    42/69

    Number of neurons in horses brains

    1,200,000,000 neurons

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    43/69

    Neuron

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    44/69

    Neuron

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    45/69

    Number of neurons in human brains

    From 19,000,000,000 to

    23,000,000,000 neurons

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    46/69

    Artificial Neural Networks

    We know that human brains consists of hugeneural network, lets study simpler artificialneural networks.

    And some day, we may be able to createartificial neural network that works the sameas the real neural network works.

    Some people thought that it will never work!They had their arguments.

    Thats why some researchers dont like neural

    networks.

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    47/69

    Neural Networks

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    48/69

    Neural Networks

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    49/69

    Backpropagation

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    50/69

    Backpropagation

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    51/69

    How to build better Neural Networks? Increase Layers?

    How would we train them?

    The network with more than 1 hidden layer is

    called a deep network

    Training a deep network is not easy using

    Backpropagation.

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    52/69

    The birth of Deep Learning Until 2006, Support

    Vector Machines always

    outperformed neural

    networks on benchmarkclassification tasks.

    At 2006, Hinton started

    the new revolution on

    machine learning: Deep

    Learning

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    53/69

    The birth of Deep Learning At 2006, Hinton proposed a greedy layer-wise

    algorithm for training a deep belief network.

    That was the first successful learning

    algorithm for a deep network. Achieved best accuracy on MNIST hand

    written digits dataset.

    Restricted Boltzmann Machines. The network works also as a generative model

    to generate data.

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    54/69

    Deep Belief Network

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    55/69

    Demo

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    56/69

    Classical Approach Train a classifier using labeled data; then test

    the classification accuracy on another testing

    data.

    Row data can be: images, videos, audio,

    natural language

    We start with extracting appropriate features.

    Hand-crafted feature extraction will differ

    according to the nature of the data.

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    57/69

    Deep Learning Approach Pre-train the network using non-labeled data-

    available.

    Fine-tune the weights using some labeled

    data. Also called: Unsupervised Feature Learning

    Deep Learning Models:

    Deep Belief Networks Stacked Autoencoders

    Convolutional Neural Networks

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    58/69

    Sparse Autoencoders

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    59/69

    What each neuron learnt?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    60/69

    What each neuron learnt?

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    61/69

    Motivating Results Unsupervised feature extraction

    outperformed hand-crafted features in set of

    popular machine recognition tasks:

    Human Activity Recognition

    Handwritten digits Recognition

    Audio classification

    ImageNet classification

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    62/69

    Biological Mapping Interesting paper at 2005: Invariant visual

    representation by single neurons in the

    human brain

    They showed that some neurons respond with

    high selectivity to, for instance, images, sound

    and name of the actress Halle Berry!

    How Many Computers to Identify a

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    63/69

    How Many Computers to Identify a

    Cat? 16,000

    Andrew NG, with Google at 2012, trained a

    very large network: on16,000 machines.

    They trained the network on random

    unlabeled videos from Youtube, without any

    guidance.

    They found that one neuron of the network

    learnt to detect human faces and another onelearnt to detect cat faces!

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    64/69

    Some neurons learnt human face!

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    65/69

    Other neurons learnt cat faces!

    How Many Computers to Identify a

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    66/69

    How Many Computers to Identify a

    Cat? 16,000

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    67/69

    Future of Artificial Intelligence

    With the increase of computational power,

    every day, scientists are able to build more

    complex neural networks

    Challenges

    Network topologies

    Learning algorithms

    Neuroscientists can lead an AI revolution if

    they came closer to how the brain works

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    68/69

    Acknowledgment

    Some slides are taken from Andrew Ng deep

    learning tutorial.

    Images used here are for educational

    purposes, any further use should return back

    to the original copyright holders.

  • 7/30/2019 Alex ACM SC Machine Learning Day [Materials] | Machine Learning& Artificial Intelligence ... Our Ultimate Dream. By Eng. Mohamad Abdelaziz Gowayed

    69/69

    Thanks You

    Questions?