Dec 2, 2013 ... CBBP, Theoretical Physics. Machine Learning. Machine Learning is the study of
computer algorithms that learn and improve through ...
Machine Learning, Module 7
Mattias Ohlsson December 2, 2013
CBBP, Theoretical Physics
Machine Learning
one of many definitions
Machine Learning is the study of computer algorithms that learn and improve through experience (data). CBBP, Theoretical Physics
Machine learning tools are (usually) constructed with the use of a labeled database Actual diagnosis
Database of labeled images
Database of medical images
Criteria
Extraction of features
Expert knowledge Construction of machine learning tool
We talk about data-driven learning CBBP, Theoretical Physics
A few examples
Classification of times series data (Electrocardiogram)
CBBP, Theoretical Physics
Sequences of letters (protein structure classification)
LSADQISTVQASFDKVKGDPVGILYAVFKADPSIMA KFTQFAGKDLESIKGTAPFETHANRIVGFFSKIIGEL PNIEADVNTFVASHKPRGVTHDQLNNFRAGFVSYM KAHTDFAGAEAAWGATLDTFFGMIFSKM
DIQMTQSPSSLSASVGDRVTITCQASQDIIKYLNWY QQTPGKAPKLLIYEASNLQAGVPSRFSGSGSGTDY TFTISSLQPEDIATYYCQQYQSLPYTFGQGTKLQIT
CBBP, Theoretical Physics
Medical images (ischemia/infarction diagnosis)
CBBP, Theoretical Physics
Medical images (finding metastatic hotspots)
CBBP, Theoretical Physics
The overall function of the machine learning tool
”Grey” box 1.2 45.4 -34.2 1.2 7.7 -30.7 21.4
CBBP, Theoretical Physics
Machine learning tool
0.12 0 No
Healthy
Outline Monday
Artificial neural networks MLP Model selection and generalization performance K-fold cross validation Bootstrap Self-Organizing Feature Map (SOFM) (Ensemble methods)
Wednesday
Ensemble methods Random forest classification Support vector machines
CBBP, Theoretical Physics
Artificial neural networks
Very simple model
Input CBBP, Theoretical Physics
∑
Output
The building blocks
Xk
y Output
Input
X0= 1
ωk
Perceptron CBBP, Theoretical Physics
The perceptron is a linear classifier
x2
o
o
o x
x
x
x
x
x x x
o
o o
x x
o
x
x x x
Class 2
x
o
o
o o o
o
x
o
o
o o
Class 1
o
o
x
o
x x
o
o
x x x
o
x x
x
o
x1
CBBP, Theoretical Physics
Build larger networks for more complex mappings
...
...
... ... ... ...
...
Multi-Layer Perceptron
CBBP, Theoretical Physics
...
...
...
CBBP, Theoretical Physics
Activation functions
CBBP, Theoretical Physics
Hidden nodes
Output nodes when classification Output nodes when function approximation
CBBP, Theoretical Physics
What can it do?
CBBP, Theoretical Physics
CBBP, Theoretical Physics
How to “train” the network?
Training data
Form an error function
CBBP, Theoretical Physics
Maximum Likelihood Principle gives
Function approximation tasks
CBBP, Theoretical Physics
Classification problems (two classes)
CBBP, Theoretical Physics
Classification problems (many classes)
CBBP, Theoretical Physics
Training is “simply minimizing the error function!
Use your favorite minimization method: Gradient descent (Backpropagation algortithm) Conjugate gradient methods (usually better than GD) Quazi Newton methods Levenberg Marquardt
Can be a complex minimization problem!!
CBBP, Theoretical Physics
How does the MLP work?
Hidden node output
hidden nodes divides the space into two parts with
as the dividing plane
CBBP, Theoretical Physics
CBBP, Theoretical Physics
CBBP, Theoretical Physics
Interpretation of outputs Start with
CBBP, Theoretical Physics
If minimization works, complex enough network and large dataset!
For classification problems, one gets
Outputs are probabilities! CBBP, Theoretical Physics
Generalization performance
Generalization performance = the performance of a trained model on new data, i.e. data that was not part of the training procedure.
The goal is to have as large as possible generalization performance
CBBP, Theoretical Physics
Overfitting can limit the generalization performance
CBBP, Theoretical Physics
How can we understand overfitting?
CBBP, Theoretical Physics
CBBP, Theoretical Physics
Overfitting can be avoided using network regularization
One common approach is weight decay
CBBP, Theoretical Physics
How to detect overfitting? Training data Data Validation data
Error
Validation error
Training error Number of iterations
CBBP, Theoretical Physics
How to estimate generalization performance? Use the idea of a validation set
All data
Training
Validation
Hold-out method
CBBP, Theoretical Physics
Use data better All data Split into K parts
V V V
... V
K-fold crossvalidation CBBP, Theoretical Physics
All data Bootstrap sample 1 Bootstrap sample 2
Bootstrap sample B
Bootstrap
CBBP, Theoretical Physics
Model selection
Error
Training error
Generalization error (estimate)
α* CBBP, Theoretical Physics
α
Estimate generalization performance with a model that needs model selection
All data Construction data
Split into K parts
Split into M parts
Test
V
Construction Test
V
Construction
CBBP, Theoretical Physics
Training
... Training
...
Construction
Training V
Construction Test
Training
Test
V
Self-Organizing Feature Maps (SOFM)
“SOFM is a kind of clustering tool”
How many clusters and where? CBBP, Theoretical Physics
Clustering can be difficult
CBBP, Theoretical Physics
K-means – A simple clustering algorithm
CBBP, Theoretical Physics
SOFM – some features
●
●
●
Artificial neural network trained using unsupervised learning (Only inputs, no targets) Input space is represented using a discretized low-dimensional (often 2-dim) space. A neighborhood function is used to preserve the topological properties of the input space (see also multidimensional scaling (MDS)).
CBBP, Theoretical Physics
Winning node A number of simple perceptrons
Define the winning node, as the one with the largest output.
CBBP, Theoretical Physics
Winning node Update the weights
CBBP, Theoretical Physics
Neighborhood functions
Note!
CBBP, Theoretical Physics
Some examples (2-D to discrete 2-D)
CBBP, Theoretical Physics
CBBP, Theoretical Physics
CBBP, Theoretical Physics
CBBP, Theoretical Physics
CBBP, Theoretical Physics