HTML document prepared by Sean Waters.

What is Machine Learning? Here is the definition given by Tom Mitchell.

Machine Learning: Any computer program that improves its performance P at some task

T through experience E.

Example: Learn to play checkers

T: play checkers & win

P: % games won in tournament

E: play against self

Example: Character recognition

T: recognize & classify hand-written characters

P: % of characters correctly classified

E: a database of hand-written characters with given classifications

Three Scenarios for Machine Learning

• Supervised Learning: each example categorized, or alternatively you receive feedback after each decision.
  • Have credit assignment problem of determining degree to which each move achieves credit or blame for final outcome.
• Reinforcement Learning: feedback is give after a sequence of actions/decisions
• Unsupervised Learning: no feedback goal is to group data into similar groups

More on Supervised Learning

Given: Training exs <x,f(x)> for some unknown function f

Find: a good approximation to f

Example Applications

• credit risk assessment

x: properties of customer and proposed purchase

f(x): approve purchase or not

• Forensic hair comparison

x: Bitmap of hair image

f(x): match or not

• Stock market prediction

x: economic indicators (S&P, Dow Jones, interest rates, …)

f(x): market will go up or down

• Steering a vehicle

x: bitmap of road surface in front of car

f(x): degrees to turn steering wheel

What types of data are there?

Feature Types:

discrete: e.g. counties in St. Louis

continuous: e.g. automobile speed

ordinal: e.g. age

relational: e.g. brother-of

How are data points x₁,x₂,… related?

independent: e.g. each is a different student

time series: e.g. x_i represents the Financial indicators of the i^th day of the year

Also need to think about independence of the features (components in x such as the

financial indicator). For example, financial indicators are not independent. Neither

is a person’s height and weight. However, a person’s shoe size and IQ are

independent.

How does data get selected?

• By a helpful teacher who selects most informative examples.

• By the learner itself who must choose between experimenting with new areas (exploration) or applying and refining knowledge of known areas (explotation). This is often referred to as exploration vs. explotation.

• By an adversary to get the worst-case behavior. Similar to worst-case asymptotic anlaysis such results can be overly pessimistic. Usually, the data requirements are smaller than given by worst-case bounds. However, such bounds are still very useful in giving guaranteed bounds.

Applications for Supervised Learning

• situations in which there is no human expert
• situations where humans can perform task but can’t describe how they do it (e.g. juggling)
• situations where the desired function, f, changes frequently
• situations where each user needs a customized function f (e.g. interest in web pages for a personalized browser, speech recognition)

Terminology

Training Ex (instance): <x,f(x)>

Target function: true function f

Hypothesis: proposed function h believed to approx f

Hypothesis space: Set of all hypotheses that can, in principle be output by the learning

algorithm.

Characterizing the Learning Process

Search Procedure (search for target in hypothesis space)

Direct Computation: solve for hypothesis directly

Local Search: start with an initial hypothesis, make small improvements until local

optimum is reached

Constructive Search: start with empty hypothesis and gradually add structure until a

local optimum is reached.

On-line vs. Batch

on-line: analyze each example and make prediction as presented

batch: collect all examples, analyze them and output hypothesis

Eager vs. Lazy

eager: analyze training data and construct an explicit hypothesis

lazy: store training data and wait until a test point is presented and then construct an

adhoc hypothesis to classify that one data point

Active vs. Passive: Can learner control the environment or run experiments vs just passively observing?

• For example, if doing stock market predictions, you really can't pick an example (i.e. the values of the indicators) and then see what happens to the market. However, if you are learning to play checkers, you can pick a board of interest and play from there. Also, you also have some control in the next board via the moves made.

Training Data vs. Test Data: Are these from same distribution?

Characterizing Hypothesis Spaces

Size: Is size fixed or variable? That is, are all hypotheses the same size? Fixed sizes

are easier to understand. Variable sizes allow more flexibility but introduce

problem of over fitting. Ex. fitting an n^th degree polynomial to a set of points.

Is hypothesis deterministic or stochastic (involves randomness)?

e.g. weather prediction – tomorrow will rain/not rain vs. tomorrow there is 80% chance

of rain

Parameters: Is each hypothesis described by a set of parameters or via a symbolic

language

Some issues in Machine Learning

• What hypothesis spaces work well?
• For these hypothesis spaces, what algorithms work well?
• How can we optimize accuracy for unseen data points?
• How does the number of training examples influence accuracy?
• What affect does noise in data have on performance?
• What are theoretical limits of learnability?
• How can prior knowledge help?
• How can systems alter their own representations?
• How can we apply machine learning in practice?