Transcript VisSupLearn178winter07

Machine Learning
ICS 178
Instructor: Max Welling
visualization & k nearest neighbors
Types of Learning
• Supervised Learning
• Labels are provided, there is a strong learning signal.
• e.g. classification, regression.
• Semi-supervised Learning.
• Only part of the data have labels.
• e.g. a child growing up.
• Reinforcement learning.
• The learning signal is a (scalar) reward and may come with a delay.
• e.g. trying to learn to play chess, a mouse in a maze.
• Unsupervised learning
• There is no direct learning signal. We are simply trying to find structure in data.
• e.g. clustering, dimensionality reduction.
Ingredients
• Data:
• what kind of data do we have?
• Prior assumptions:
• what do we know a priori about the problem?
• Representation:
• How do we represent the data?
• Model / Hypothesis space:
• What hypotheses are we willing to entertain to explain the data?
• Feedback / learning signal:
• what kind of learning signal do we have (delayed, labels)?
• Learning algorithm:
• How do we update the model (or set of hypothesis) from feedback?
• Evaluation:
• How well did we do, should we change the model?
Data Preprocessing
• Before you start modeling the data, you want to have a look at it to get a “feel”.
• What are the “modalities” of the data: e.g.
• Netflix: users and movies
• Text: words-tokens and documents
• Video: pixels, frames, color-index (R,G,B)
• What is the domain?
• Netflix: rating-values [1,2,3,4,5,?]
• Text: # times a word appears: [0,1,2,3,...]
• Video: brightness value: [0,..,255] or real-valued.
• Are there missing data-entries?
• Are there outliers in the data? (perhaps a typo?)
Data Preprocessing
• Often it is a good idea to compute the mean and variance of the data.
•
1
E[ X i ] 
N
N
X
n 1
mean
in
1
VAR[ X i ] 
N
N
(X
n 1
in
variance
 E[ X i ]) 2
STD[ X i ]  Var[ X i ]
standard deviation
• Mean gives you a sense of location, Variance/STD a sense of scale.
• Better even is to histogram the data:
Tricky issue: how do you choose the bin-size:
too small: you see noise, too big: it’s one clump.
Preprocessing
• For netflix you can histogram this for both modalities:
• The rating distribution over users for a movie.
• The rating distribution over movies for a user.
• The rating distribution over users for all movies jointly.
• The rating distribution over all movies for all users jointly.
user-variance
• You can compute properties and plot them against each other.
For example:
• Compute the the user-specific mean variance over movies and plot
a scatter plot:
every dot is a different user
user-mean
Scatter-Plots
This shows all the
2-D projections of the
“Iris data”.
Color indicates the class
of iris.
How many attributes
do we have for Iris?
3-D visualization
contour plot
meshgrid plot
Embeddings
• Every red dot represents an image.
• An image has +/- 1000 pixels
• Each image is projected to a 2-D
space
• Projections are such that similar
images are projected to similar
locations in the 2-D embedding.
• This gives us an idea how the data
is organized.
These plots are produced by “local linear embedding”
http://www.cs.toronto.edu/~roweis/lle/
Embeddings
Visualization by Clustering
By performing a clustering of the data and looking at the cluster-prototypes
you can get an idea of the type of data.
Preprocessing
• Often it is useful to “standardize” (or “whiten”) the data before you start modeling.
• The idea is to remove the mean and the variance so that your algorithm can
focus on more sophisticated (higher order) structure.
1) X in  X in  E[ X i ]
2) X in 
X in
STD[ X i ]
In that order!
Be Creative!
WEKA DEMO