visual learning with navigation as an example
DESCRIPTION
Visual Learning with Navigation as an Example. Dr Juyang Weeng Dr Shaoyun Chen Michigan Sate University. Model Based Methods. PROS Efficient for predictable cases Easier to understand Computationally inexpensive CONS Non generic Not able to deal with every possible case - PowerPoint PPT PresentationTRANSCRIPT
Visual Learning with Navigation as an Example
Dr Juyang WeengDr Shaoyun Chen
Michigan Sate University
PROS Efficient for predictable cases Easier to understand Computationally inexpensive CONS Non generic Not able to deal with every possible case Potentially huge number of exhaustive
cases.
Model Based Methods
Example of Model based learning
[1]
Automatically learn the model
Xt input image in rc dimensional space(S)Yt+1 control signal in space CThe image needs to be vectorized.
GOAL :Approximate the function f
MODEL FREE METHODS
Yt+1=f(Xt)
Each leaf node represents sample(X,Y)
Each node represents a set of data points with increased similarity
One of the central ideas in Shoslif’s approach
Given X find f(X) at the corresponding leaf node after traversal.
Recursion Partition Tree
Building a Regression Partition Tree Take the sample space S. Divide the space into b cells. Each a child of
the root. The analysis performs automatic derivation
of features(discussed later). Continue to do this until the leaf nodes have
a single data point or many data points with virtually the same Y.
Learning Phase
How to construct the RPTLearning Phase
6 7
12 3
4 5
8 9
12
3
45
6 7
8
9
Input X’ Output Y control signal Recursively analyze the centre of each node If it is close to the input then proceed in that
direction till you reach the leaf node . Use the corresponding Control signal Use top k paths to find the top k nearest
centers.
Performance phase
Feature Selection :Select features from a set of human defined features.
Feature Extraction: extrapolates selected features from images
Feature Derivation : derives features from high dimensional vector inputs
Using Principal Component Analysis recursively partitions the space S into a subspace S’ where the training samples lie.
Automatic Feature Derivation
Computes the principal component vectors .◦ V1,V2,V3,V4…..VN
MEF : Most Expressive Features They explain the variation in the sample set The hyper plane that has V1 as a normal an
that passes through the centroid of the samples forms a partition.
The samples on one side fall onto on side of the tree and vice versa.
PCA
PCA v/s LDA
[1]
PCA LDA
We can do better with class information. MDF :Most discriminating feature Similar to PCA This method is cuts more along the class
boundaries.
Differences MEF: samples spread out widely, and the samples
of different classes tend to mix together. MDF: samples are clustered more tightly, and the
samples from different classes are farther apart.
LDA
Using a model similar to Markov chain model
St State at time t
At time t, the system is at state St and observes image Xt.
Control vector Yt+1 and enters the next state St+1.(St+1, Yt+1) = f (St, Xt)
Using States
The Observation driven Markov Model[1]
[1]
A special state A (ambiguous) indicates that local visual attention is needed.
Eg. trainer defined this state for a segment right before a turn.
If the image area that revealed the visual difference between different turn types was mainly in a small part of the scene.
A directs the system to look at such landmarks through a prespecified image sub window so that the system
issues the correct steering action before it is too late.
Dealing with local attention
Batch learning : All the training data are available at the time the system learns.
Incremental learning : Training samples are available only one at a time.
Discard once you have used them Memory requires to store the image only
once. Similar images discarded
Incremental Learning
The Learning Process
Step 1•Query the current RPT.
Step 2•If the difference between the current RPT’s output
and the desired control signal >prespecified tolerance.
•Go to Step 3 Else Goto Step 1
Step 3 •Shoslif learns the current sample to update the RPT.
Compared Shoslif with feed forward neural networks and radial basis function networks for approximating stateless appearance-based navigation systems.
Shoslif did significantly better than both methods.
Extension to face detection, speech recognition and vision-based robot arm action learning.
Shoslif versus other methods
Shoslif performs better in benign scenes.
The state based method allows more flexibility
However still need to specify that many states for different environment types.
Conclusion
1. Dr.Juyang Weng & Dr. Shaouyun Chen “Visual Learning with Navigation as an Example” .Published in IEE September/October 2000.
References
Questions