Lecture 1: Introduction to the Course EECS 545: Machine Learning - - PDF document

1

Lecture 1: Introduction to the Course

EECS 545: Machine Learning Benjamin Kuipers

EECS 545: Machine Learning

• Professor Benjamin Kuipers

– 647-6887, kuipers@umich.edu – Office hours: TTh 2:00 - 3:00, CSE 3741

• GSI: Gyemin Lee

– gyemin@eecs.umich.edu – Office hours: MW 1:00 - 2:30, EECS 2420

2

Who is Benjamin Kuipers?

• This is my first semester at Michigan.

– University of Texas at Austin, 1985 - 2008 – MIT, 1984 - 1985 – Tufts University, 1978 - 1984 – MIT, 1972 - 1978; PhD, 1977 – Harvard University, 1970 - 1972 – Swarthmore College, 1966 - 1970; BA, 1970. – Ann Arbor High School, 1963 - 1966

Who is Benjamin Kuipers?

• Research on representation and use of

commonsense knowledge.

– Modeling the cognitive map of large-scale space

• Critical role of the topological map

– Robot exploration and mapping

• Concept of distinctive state

– Qualitative reasoning about physical systems

• QSIM algorithm for qualitative simulation

– Learning from uninterpreted sensors and effectors

• Bootstrap learning from pixels to high-level concepts

3

What is Machine Learning?

• Learning, done by a machine!

– That is, learning well enough understood that we can formally describe the process and implement it as an algorithm.

• It’s a branch of artificial intelligence.

– The problem of understanding the mind as a computation, and how a physical system can have a mind.

What is Learning?

• Learning facts about the world.

– French explorers enter Michigan in 1621. – University of Michigan established 1817. – Michigan becomes a state in 1837.

• Collecting facts is relatively superficial,

but still raises significant problems.

– Q: What is the knowledge representation?

4

What is Learning?

• Learning facts about the world.

– French explorers enter Michigan in 1621. – University of Michigan established 1817. – Michigan becomes a state in 1837.

• Spanish explorers enter Texas in 1519.
• Texas independence from Mexico 1836.
• Texas becomes a state in 1845.
• University of Texas established 1883.
• Collecting facts is relatively superficial,

but still raises significant problems.

– Q: What is the knowledge representation?

What is Learning?

• Learning to classify observations.

– Classifying observed plants or animals – Diagnosing diseases from symptoms and tests – Identifying faces in images

• Supervised learning: given a training set

– {feature vectors plus classifications} – learn to classify new instances – Q: What are the best features?

5

What is Learning?

• Learning skilled action.

– Learning to balance, and walk, then run. – Learning to ride a bicycle.

• Reinforcement learning:

– given a reward signal – learn sequences of actions to maximize discounted future reward. – Q: Where does the reward signal come from?

What is Learning?

• Learning useful categories and features.

– Dogs are more similar to other dogs than to cats. – Retrievers resemble other retrievers. – Chihuahuas resemble other chihuahuas.

• Unsupervised learning:

– Clustering by similarity defines categories. – Features that discriminate well are useful. – Dimensionality reduction to a few features. – Q: Which categories and features are useful?

6

How does a baby (human or robot) get knowledge of its own?

• The baby, assailed by eyes, ears, nose, skin,

and entrails at once, feels it all as one great blooming, buzzing confusion …

– [William James, 1890]

• How does it get high-level concepts?

– Places, paths, objects, actions, plans, etc. – Foundational domains: space, time, . . .

Our Gedankenexperiment

• Consider a baby robot, a learning agent, born

with uninterpreted sensors and effectors

– We pretend that evolutionary learning is done by the individual, not by the species.

• The baby has only pixel-level experience:

– Disorganized collection of sensor elements – Incremental motor signals

• How does it learn object-level concepts?

– Places, paths, objects, actions, etc. – The macro-scale components of adult knowledge

7

Here’s the platform I want!

• But here’s what we have …

Lassie “sees” the world with a Laser Rangefinder

• 180 ranges over

180° planar field of view

• About 13” above

the ground plane

• 10-12 scans per

second

8

Laser Rangefinder Image

• 180 narrow beams at 1º intervals.

Disorganized Sensor: 180 “Pixels”

9

Structured Sensor Array The Egocentric Range Image

10

The World-Centered Range Image The World-Centered Range Image

11

Image History in 2D Histogram Statistical Learning Methods Used

• Correlation (time-series and histograms)
• k-means and agglomerative clustering
• Multidimensional scaling
• Dimensionality reduction (PCA, Isomap)
• Sensory flow
• Image matching (ICP)
• Markov localization (max likelihood pose)
• . . .

12

Objects as Explanations

• A static model explains most observations.

– So focus on the discrepancies

• Cluster in space; Track over time
• Merge images to make shape models

– Modayil & Kuipers [2004, 2006, 2007, 2008].

Identify Dynamic Sensor Returns

13

Clustering into Objects Track Objects over Time

14

Describe the Scene

• Describe the scene

in terms of:

– Static world – Robot’s own pose – Object in a fixed position – Object and trajectory

• Individual objects

Learning Object Shapes

• Merge range scans

to get shape models

• Cluster shapes to

get object categories

15

Learning Object Categories

• Clustering shapes by perceptual features

Learn about Actions

• Learn actions to affect objects. Learn:

– Qualitative description of effect – Bounds on prerequisite state – Control law to perform the action

• For a mobile robot that can move and push:

– Move to desired point in nearby space. – Turn to face object. – Push (Move, to get object to move also)

16

Learn about Actions

• Learn action properties. Simple planning.

From Pixels to Concepts

• Learn high-level concepts by identifying

structure in the pixel stream.

– Space is learned as a minimal explanation for sensory correlations. – Objects are learned as a minimal explanation for discrepancies from fixed-value model – Actions are learned as minimal descriptions of motions interacting with objects – Plans combine actions to achieve goals.

• What are the methods for doing this?

17

EECS 545: Machine Learning

• The purpose is to learn these methods.

– Statistical learning from Bishop. – Reinforcement learning from Sutton & Barto.

• Reading and lectures cover the Theory part.

– Homework problems are an important part. – Do them, and grade them, in groups. – Presentations to the class.

• Programming projects cover the Practice part.

– Each person submits their own projects. – Help each other learn and do the projects.

Work Together

• Help each other learn the material.

– Read and discuss it with each other. – Catch and correct each other’s mistakes. – Don’t let someone avoid learning. – Think like a teacher!

• We will form groups.

– Do your homework (blue or black pencil or pen). – Correct someone else’s homework (red pen). – Hand it in (for a completion grade).

18

Programming Projects

• Applying machine learning to a problem:

– In Theory there’s no difference between theory and practice. – But in Practice, there is.

• We’re going to learn to find and describe
• bjects in a stream of video input.

– Individual projects. – Help each other succeed.

Understanding Video Input

• Video: a high-density stream of pixels.

– The viewer imposes an interpretation. – How? (exploit redundancy)

• Build a model of static background (#1).
• Individuate and track foreground objects (#2).
• Stabilize variations in static background (#3).
• Build models of foreground objects (#4).

19

Traffic Video; Fixed Camera Static Background

20

Foreground “Objects” Stabilize Hand-Held Camera

21

Open Issues

• Collect your own videos and test your methods.
• More complex pixel models for #1, to handle

waving branches, etc.

– Read Stauffer & Grimson, PAMI, 2000.

• How can we learn that shadows in #2 are not parts
• f objects?
• #3 is essentially a SLAM problem (Simultaneous

Localization And Mapping).

– How can we focus attention on informative features?

• Need a new video for #4, to make object modeling

a 2D problem (tractable), rather than a 3D problem (too hard).

Plan for Class Sessions

• Each class is 80 minutes: 12:10 - 1:30 pm

– 30-40 minutes for my lecture – 20-30 minutes for student presentation

• Need volunteers for next Tuesday and Thursday!

– 20-30 minutes to discuss projects

• Attend every class session. And exams:

– Mid-term: February 19 (12:00 - 1:30 pm). – Final exam: April 30 (1:30 - 3:30 pm).

22

Now Form Groups!

• Each group should have 2-4 students.

– Make sure your schedules are compatible!

• Submit a sheet for each group.

– Names, phones, and emails for all members. – Keep copies for yourselves.

• Volunteers to present problem solutions

next Tuesday and Thursday.

– (These should be easier.)