Machine Learning in Five Minutes

1
Machine Learning in Five Minutes
Spring 2003

• Charles Isbell
• Jeff Pierce

2
Agenda

• A case study
• An overview
• Resources

3
Machine Learning

• Adaptation of an artifacts behavior over time
• Think about the data not the programming
• Think about the features
• Semantics might just be the things you do with
  something
• Pattern recognition
• Statistics
• Probability theory
• information theory
• Theorems Algorithms
• Less about engineering, a lot less about
  scalability
• Can be about analysis, visualization, etc

4
A Case Study

• Task Management
• What can I do next?
• Is my schedule reasonable?
• Hand-coding rules vs Learning
• Representation identifying the stand-in features
• Finding those features
• Unifying representations
• Bringing it all together

5
Machine Learning Methodologies

• Supervised Learning
• Labeled examples
• Function approximation
• Unsupervised Learning
• Description as density estimation
• Clustering of unlabeled data
• Dimensionality Reduction
• Reinforcement Learning
• Labeled sequences of events
• A method for learning action choice (from delayed
  labeling)

6
Supervised Learning

• Classification
• Function Approximation
• Give me lots of pairs of (x,y) and Ill learn
  yf(x)
• Neural networks
• Decision trees
• SVMs
• Time series

7
Unsupervised Learning

• Description
• Clustering (classification?)
• Density Estimation
• Time Series
• Dimensionality Reduction
• PCA, ICA
• K-clustering
• HMMs, K-order MMs
• Graphical Models
• Signal Separation

8
Reinforcement Learning

• The Approach
• States (may be only partially observable)
• Actions (things one can do)
• Transitions (how an action changes state)
• Rewards (signals from the environment)
• Policy maximizes long-term expected reward
• Well-Defined and Understood
• History of Success
• TD-Backgammon, cobot, Robosoccer, Robotics, Food
  Processing
• Temporal Difference, Q-Learning, Policy Iteration

9
Slogans

• Experience into expertise
• Focus on application-independent approaches for
  improving behavior via learning experience is
  the critical resource
• Start in the middle and you're stuck in the
  middle
• A system must ultimately derive its own
  representations from data from scratch or we'll
  be forever maintaining them
• Data, not data structure
• Traditional knowledge representations are
  brittle, limited, and seemingly impossible to
  derive from data
• Seeing is "seeing as"
• Perception and understanding are about making
  connections to past contexts via experience-tuned
  similarity metrics

10

• Program a computer to fish and you have taught it
  one skill.
• Program a computer to learn and you have taught
  it everything.

11
Resources

• Association Agent
• Ishmail
• HMM
• stuff on the web