Getting a Machine

parathyroidsanchovyAI and Robotics

Nov 17, 2013 (3 years and 4 months ago)

49 views

Getting a Machine

to Learn

Danny Silver, PhD. CIM.

March 22, 2006

What is Learning?


The process of
acquiring knowledge
or skill through study,
experience or
teaching


Fundamental to
success and survival


What is Learning?


Inductive inference/modeling


Developing a general model/hypothesis from
examples


Face Recognition …


Happy Face

recognition, that is!


It’s like …
Fitting a curve to data



Also considered modeling the data


Statistical modeling


What is Learning?


Requires an inductive bias

= a heuristic beyond the data



Do you know any inductive biases?



How do you choose which to use?

Inductive Biases


Universal heuristics
-

Occam’s Razor


Knowledge of intended use


Medical diagnosis


Knowledge of the source
-

Teacher


Knowledge of the task domain


Analogy with previously learned tasks



Inductive Bias and

Knowledge Transfer

ASH ST

THI RD

SEC OND

ELM ST

FIR ST

PINE ST

OAK ST

Inductive bias depends upon:



Having prior knowledge



Selection of most related
knowledge

Human learners use Inductive Bias

What is Machine Learning?


The study of how to build computer
programs that:


Improve with experience


Generalize from examples


Self
-
program, to some extent




What is Machine Learning?


Function approximation




(curve fitting)



Classification
(concept learning, pattern
recognition)

x1

x2

A

B

f(x)

x

What is Machine Learning?

Basic Framework for Inductive Learning

Inductive

Learning System

Environment

Training

Examples

Testing

Examples

Induced Model

or Hypothesis

h

Output Classification

(x, f(x))

(x, h(x))

h(x) = f(x)?

A problem of
representation and

search
for the best hypothesis, h(x).

~

Why Study Machine Learning?


Computer Science



theory of computation, new
algorithms


Math

-

advances in statistics, information theory


Psychology



as models for human learning, knowledge
retention and transfer


Biology



how does a nervous system learn


Philosophy



epistemology, knowledge acquisition


Application



new knowledge extracted from data,
solutions to unsolved problems


History of Machine Learning

[Patterson, D., Artificial Neural Networks: Theory and Applications, 1996,
Figure 1.10 p13]

Classes of ML Methods


Supervised



Develops models that predict the value of
one variable from one or more others:


Artifical neural networks, inductive decision trees, genetic
algorithms, k
-
nearest neighbour, Bayesian Networks, support
vectors machines


Unsupervised



Generates groups or clusters of data
that share similar features


ANN, Self
-
organizing feature maps


Reinforcement Learning



Develops models from the
results of a final outcome; like winning the game


TD
-
learning, Q
-
learning



How do we get a Machine to Learn?


Problem: We wish to learn to classifying two people
(A and B) based on their keyboard typing.


Approach:


Acquire lots of typing examples from each person


Extract relevant features ??
-

representation!


Transform feature representation as needed


Use an algorithm to fit a model to the data
-

search!


Test the model on an independent set of examples of typing
from each person

Classification

Mistakes

Typing Speed

A

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

B

B

B

B

B

B

B

B

B

Logistic Regression

Y

Z=f(M,T)

0

1

Classification

A

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

B

B

B

B

B

B

B

B

B

Artificial Neural Network

A

Mistakes

Typing Speed

M

T

Y

Classification

A

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

B

B

B

B

B

B

B

B

B

Inductive Decision Tree

A

A

Mistakes

Typing Speed

M?

T?

T?

Root

Leaf

A

B

Blood Pressure Example

Classification

A

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

B

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

A

B

B

B

B

B

B

B

B

B

k

Nearest Neighbour

A

Mistakes

Typing Speed

Classification is based

on majority vote of

nearest neighbours.

How do we get a Machine to Learn?



Demo
-

Typist Identification


Application: user validation
-

BioPassword

How can we be certain our

Machine has Learned?


Generalization accuracy can be guaranteed
for a specified confidence level given
sufficient number of examples


Models can be validated for accuracy by using
a previously unseen test set of examples



Sound familiar?
Exams measure a students
generalized knowledge of subject.

Enter: Learning Theory

P
robably
A
pproximately
C
orrect (PAC)

theory
of learning
[Leslie Valiant, 1984]


Premise:

Generalization must be based on test
examples drawn from the same input space and
with the same probability distribution


Poses questions such as:


How many examples are needed for good generalization?


How long will it take to create a good model?


Answers depend on:


Complexity of the actual function


The desired level of accuracy of the model (75%)


The desired confidence in finding a model (19/20=95%)

Linear and Non
-
Linear Problems


Linear Problems


Linear functions


Linearly separable



classifications


Non
-
linear Problems


Non
-
linear functions


Not linearly separable


classifications

x1

x2

A

B

x1

x2

f(x)

x

A

B

B

f(x)

Application Areas


Data mining
-

predictive and descriptive models


Web mining


information filtering


User Modeling


adaptive systems


Intelligent Agents


cooperative multi
-
agent
systems


Robotics


image recognition


Application Areas

Data Mining Applications:


Science and medicine:

prediction, diagnosis, pattern recognition


Manufacturing:

process modeling and analysis


Marketing and Sales:

targetted marketing, segmentation


Finance:

portfolio trading, investment support


Banking & Insurance:

credit and policy approval


Security:

bomb, iceberg, fraud detection


Engineering:

dynamic load shedding, pattern recognition


The Future


Machine Lifelong Learning
-

Inductive Bias
Revisited


Robotics: from Martian Rovers to the 101
Highway


User Modeling for Complex HCI and the
Semantic Web

Machine Lifelong Learning (ML
3
)


Concerned with persistent and cumulative nature of
learning
[Thrun’97]


Learner faces a sequence of learning tasks over time


Investigates methods of improving effectiveness and
efficiency of learning over the sequence


Our focus: retention of prior task knowledge and its
use as a source of inductive bias



Research Software


Robotics: OASIS for NASA Rovers

Onboard Autonomous Science Investigation System



The capacity of the rover to collect data is surpassing
bandwidth to transmit


Opportunity to increase mission science return by
selecting the data with the highest science interest


OASIS evaluates geologic data gathered by the rover


ML component identifies potential science
opportunities


ML also prioritizes the data for transmission



Stanford Racing Team's leader Sebastian Thrun holds a
$2
-
million dollar check as he catches a ride on top of
Stanley No. 03, a tricked
-
out Volkswagen Touareg R5,
after his team was declared the official winner of the
DARPA Grand Challenge 2005 in Primm, Nevada.

Source: Associated Press


Saturday, Oct 8, 2005

DARPA Grand Challenge 2005

DARPA Grand Challenge 2005


Stanley the VW Touareg, designed by Stanford University,
zipped through the 132
-
mile Mojave Desert course in six hours
and 53 minutes Saturday, using only its computer brain and
sensors to navigate rough and twisting desert and mountain
trails.


According to Thrun and Mike Montemerlo, a postdoc who was
the software guru for the Stanford team, this robot had the
ability to learn about the road. Its sensors gathered information
about what was underneath its front bumper and used that
knowledge to figure out what was road and what was not road
for hundreds of feet ahead. Also, when it came to figuring out
what should be avoided and what could be ignored, Stanley was
trained to emulate the behavior of human drivers.

User Modeling and Adaptive Systems


Expertise: Machine Learning


Sub
-
area of artificial intelligence


Development of predictive models from examples


Application to User Modeling and Identification

User

User Interface

Application

Software


Learning

System

User

Model

Explicit data



preferences,
chosen options

Implicit data

-

keystroke
and mouse click traces

User Modeling

Intelligent Web Filters

Form Field Ordering

and Completion

Smart Email

Handheld Fashion Consultant

User Identification

Key Stroke Biometrics

Smart Navigator

Handwriting ID

Eye
-
tracking Biometrics

THE END


danny.silver@acadiau.ca


http://osiris.sunderland.ac.uk/cbowww/AI/TEXTS/ML2/index.htm

Fuzzy Sets vs Probability

Fuzzy sets are based on vague definitions of sets, not randomness.

To illustrate the difference, consider this scenario: Bob is in a house
with two adjacent rooms: the kitchen and the dining room. In
many cases, Bob's status within the set of things "in the kitchen"
is completely plain: he's either "in the kitchen" or "not in the
kitchen". What about when Bob stands in the doorway? He may
be considered "partially in the kitchen". Quantifying this partial
state yields a fuzzy set membership. With only his little toe in the
dining room, we might say Bob is 0.99 "in the kitchen", for
instance. No event (like a coin toss) will resolve Bob to being
completely "in the kitchen" or "not in the kitchen", as long as
he's standing in that doorway.

Source:
http://en.wikipedia.org/wiki/Fuzzy_logic