Copyright R. Weber Neural Networks INFO 629 Dr. R. Weber.
-
Upload
mavis-sherilyn-fitzgerald -
Category
Documents
-
view
225 -
download
1
Transcript of Copyright R. Weber Neural Networks INFO 629 Dr. R. Weber.
Copy
right
R. W
eber
Neural Networks
INFO 629 Dr. R. Weber
Copy
right
R. W
eber
~= 2nd-5th week
training vision
the evidence
Copy
right
R. W
eber
the evidence
~= 2nd-5th week
training vision
10
Copy
right
R. W
eber
the evidence
~= 2nd-5th week
training vision
10
Copy
right
R. W
eber
the evidence
~= 2nd-5th week
training vision
Copy
right
R. W
eber
NN: model of brains
input output
neuronssynapses
electric transmissions:
Copy
right
R. W
eber
Elements
• input nodes• output nodes• links• weights
Copy
right
R. W
eber
terminology
• input and output nodes (or units) connected by links
• each link has a numeric weight
• weights store information
• networks are trained on training sets (examples) and after are tested on test sets to assess networks’ accuracy
• learning/training takes place as weights are updated to reflect the input/output behavior
Copy
right
R. W
eber
The trainingOutput=Step(w f )
learning takes place as weights are updated to reflect the input/output behavior
=> mammal (1)=> bird (0)
0 1 1
4 legs flies eggs
i=1
i=2
i=3
j=1 j=2 j=3
ij
0 0 0
0 0 0
0 0 0
1 0 0
1 0 0
1 0 0
1 0 0
1 0 0
1 1 1
1 0 0
1 1 1
1 1 1
Goal minimize error between representation of the expected and actual outcome
20
ij
Copy
right
R. W
eber
The concept i
=> mammal
=> bird0 1 1
4 legs flylayeggs
1 0 0
1 Yes, 0 No
=> mammal1 1 0
Copy
right
R. W
eber
The concept ii
=> mammal
=> bird0 1 1
4 legs flylayeggs
1 0 0
=> mammal1 1 0
1 Yes, 0 No
Copy
right
R. W
eber
The learning i
=> mammal
=> bird0 1 1
4 legs flylayeggs
1 0 0
=> mammal1 1 0
0.5 0.5 0.5
1 Yes, 0 No
Copy
right
R. W
eber
=> mammal
=> bird0 1 1
4 legs flylayeggs
1 0 0
=> mammal1 1 0
0*0.5+1*0.5+1*0.5= 1
1*0.5+0*0.5+0*0.5= 0.5
1*0.5+1*0.5+0*0.5= 1Goal is to have weights that recognize different representations of mammals and birds as such
0.5 0.5 0.5
The learning ii
Copy
right
R. W
eber
=> mammal
=> bird0 1 1
4 legs flylayeggs
1 0 0
=> mammal1 1 0
0*0.5+1*0.5+1*0.5= 1
1*0.5+0*0.5+0*0.5= 0.5
1*0.5+1*0.5+0*0.5= 1Suppose we want bird to be greater 0.5 and mammal to be equal or less than 0.5
0.5 0.5 0.5
The learning iii
Copy
right
R. W
eber
=> mammal
=> bird0 1 1
4 legs flylayeggs
1 0 0
=> mammal1 1 0
0*0.25+1*0.25+1*0.5= 0.75
1*0.25+0*0.25+0*0.5= 0.25
1*0.25+1*0.25+0*0.5= 0.5Suppose we want bird to be greater 0.5 and mammal to be equal or less than 0.5
0.25 0.25 0.5
The learning iv
Copy
right
R. W
eber
NN demo…..
Copy
right
R. W
eber
Characteristics
• NN implement inductive learning algorithms (through generalization) therefore, it requires several training examples to learn;
• NN do not provide an explanation or a rule telling why the task performed the way it was;
• uses data rather than explicit knowledge; • Classification (pattern recognition), clustering,
diagnosis, optimization, forecasting (prediction), modeling, reconstruction, routing;
Copy
right
R. W
eber
NN Parameters• Number of hidden layers,• Randomly initialization of the weights• Error, threshold• Adjustment rate for weights fixed or variable
• Optimal parameters are determined empiricallyA. Abraham and B. Nath, “Hybrid Heuristics for
Optimal Design of Neural Nets,” in Proceedings of the Third International Conference on Recent Advances in Soft Computing, R. John and R. Birkenhead, Eds. Germany: Springer Verlag, 2000, pp. 15-22.
Copy
right
R. W
eber
Network Topology: Feedforward & Feedback
• Feedforward– connections between units do not form cycles– produce a response to an input quickly– can be trained using a wide variety of efficient
conventional numerical methods
• Feedback (recurrent NN)– connections between units form cycles– Takes a long time before it produces a response– more difficult to train than feedforward
Copy
right
R. W
eber
• Feedforward– Linear, Perceptron, Adaline , Higher Order,
Functional Link, MLP: Multilayer perceptron, Backpropagation, Cascade Correlationm, Quickprop, RPROP , Radial Basis Function networks, OLS: Orthogonal Least Squares, CMAC: Cerebellar Model Articulation Controller, Classification only, LVQ: Learning Vector Quantization, Kohonen, PNN: Probabilistic Neural Network, GNN: General Regression Neural Network
Copy
right
R. W
eber
• Feedback – BAM: Bidirectional Associative Memory,
Boltzman Machine, Recurrent time series, Backpropagation through time , FIR: Finite Impulse Response, Real-time recurrent network, Recurrent backpropagation, TDNN: Time Delay NN
Copy
right
R. W
eber
Where are NN applicable?
• Where they can form a model from training data alone;
• When there may be an algorithm, but it is not known, or has too many variables;
• There are enough examples available• It is easier to let the network learn from
examples• Other inductive learning methods may not be
as accurate
Copy
right
R. W
eber
Applications (i)• predict movement of stocks, currencies, etc.,
from previous data;• to recognize signatures made (e.g. in a bank)
with those stored;• Classify medical imaging e.g., ECG signal, X
Rays, MRIs• classify the state of aircraft engines (by
monitoring vibration levels and sound, early warning of engine problems can be given; British Rail have been testing an application to diagnose diesel engines;
Copy
right
R. W
eber
Applications (ii)• Pronunciation (rules with many exceptions);• Handwritten character recognition
(network w/ 200,000 is impossible to train, final 9,760 weights, used 7300 examples to train and 2,000 to test, 99% accuracy)
• ATR automated target recognition, distinguish threatening from non threatening targets;
• Learn brain patterns to control and activate limbs as in the “Rats control a robot by thought alone” article
• Credit assignment
Copy
right
R. W
eber
Applications (iii)• Optimization (max,min) and routing (min
distances) problems • Modeling e.g., create model of input
vs.output analysis of software programs to work as an oracle of predicted output
• Reconstruction to produce clean versions of noisy patterns by matching the closest training pattern to input pattern
• Clustering
Copy
right
R. W
eber
CMU Driving ALVINNhttp://www.ri.cmu.edu/projects/project_160.html
• learns from human drivers how to steer a vehicle along a single lane on a highway
• ALVINN is implemented in two vehicles equipped with computer-controlled steering, acceleration, and braking
• cars can reach 70 m/h with ALVINN• programs that consider all the problem
environment reach 4 m/h only
Copy
right
R. W
eberWhy using NN for the driving task?
• there is no good theory of driving, but it is easy to collect training samples
• training data is obtained with a human* driving the vehicle–5min training, 10 min algorithm runs
• driving is continuous and noisy• almost all features contribute with useful
information*humans are not very good generators of training instances when they behave too regularly without making mistakes
Copy
right
R. W
eber
• INPUT:video camera generates array of 30x32 grid of input nodes
•OUTPUT: 30 nodes layer corresponding to steering direction
•vehicle steers to the direction of the layer with highest activation
the neural network
Copy
right
R. W
eber
Genetic algorithms (i)
• learn by experimentation• based on human genetics, it originates new
solutions • representational restrictions• good to improve quality of other methods
e.g., search algorithms, CBR• evolutionary algorithms (broader)
Copy
right
R. W
eber
Genetic algorithms (ii)
• requires an evaluation function to guide the process• population of genomes represent possible solutions• operations are applied over these genomes• operations can be mutation, crossover• operations produce new offspring• an evaluation function tests how fit an offspring is • the fittest will survive to mate again
Copy
right
R. W
eber
Genetic Algorithms (iii)
• http://ai.bpa.arizona.edu/~mramsey/ga.html You can change parameters
• http://www.rennard.org/alife/english/gavgb.html Steven Thompson presented