Derivation of Fischer's Linear Discriminant
===============
Main article: [Derivation of Fisher's Linear Discriminant]
The derivation was completed in this lecture.
Recall from last lecture
=====
Last time, we considered
$ J(\vec{w}) = \frac{\vec{w}^t S_B \vec{w}}{\vec{w}^t S_W \vec{w}} $
|jinha_Jw|
.. |jinha_Jw| image:: tex
:alt: tex: J(\vec{w}) = \frac{\vec{w}^t S_B \vec{w}}{\vec{w}^t S_W \vec{w}}
which is explicit function of $ \vec{w} $
.. |jinha_vecw| image:: tex
:alt: tex: \vec{w}
One can do this because numerator of $ J(\vec{w}) $ can be written as
$ \mid \tilde m_1 - \tilde m_2 \mid^2 = \mid w \cdot (m_1 - m_2) \mid^2 = w^t (m_1 - m_2) (m_1^t - m_2^t) w $
$ \rightarrow S_B = (m_1 - m_2) (m_1^t - m_2^t) $
|jinha_top|
.. |jinha_top| image:: tex
:alt: tex: \mid \tilde m_1 - \tilde m_2 \mid^2 = \mid w \cdot (m_1 - m_2) \mid^2 = w^t (m_1 - m_2) (m_1^t - m_2^t) w
|jinha_sb|
.. |jinha_sb| image:: tex
:alt: tex: \rightarrow S_B = (m_1 - m_2) (m_1^t - m_2^t)
In a same way, denominator can be written as
$ \tilde s_1^2 + \tilde s_2^2 = \sum_{y_i \in class \ i} (w \cdot y_i - \tilde m_1)^2 = \sum w^t (y_i - m_i)(y_i^t - m_i^t) w $
$ = w^t \left[ \sum (y_i - m_i)(y_i^t - m_i^t) \right] w $
$ \rightarrow S_W = \sum_{y_i \in class \ i} (y_i - m_i)(y_i^t - m_i^t) $
Fisher Linear Discriminant
=======
.. |omega_0| image:: tex
:alt: tex: \omega_0
.. |generalized_evalue_prob| image:: tex
:alt: tex: S_B\omega_0=\lambda S_W\omega_0
.. |standard_evalue_prob| image:: tex
:alt: tex: {S_W}^{-1}S_B\omega_0=\lambda\omega_0
- Maximizing J as defined in previous class
It is a known result that J is maximum at |omega_0| such that |generalized_evalue_prob|. This is the "Generalized eigen-value problem.
.. |notsingular| image:: tex
:alt: tex: |S_W|\neq 0
Note that if |notsingular|, then |standard_evalue_prob|. It can be written as the "Standard eigenvalue problem". The only difficulty (which is a big difficulty when the feature space dimension is large) is that matrix inversion is very unstable.
.. |cst_meandifference| image:: tex
:alt: tex: S_B\omega_0=(\vec{m1}-\vec{m2})(\vec{m1}-\vec{m2})^T\omega_0=cst.(\vec{m1}-\vec{m2})
.. |reduced_prob| image:: tex
:alt: tex: {S_W}^{-1}cst.(\vec{m1}-\vec{m2})=\lambda\omega_0
.. |value_omega_0| image:: tex
:alt: tex: \omega_0={S_W}^{-1}(\vec{m1}-\vec{m2})
Observe that |cst_meandifference|. Therefore the standard eigenvalue problem as presented above becomes |reduced_prob|. From this equation, value of |omega_0| can easily be obtained, as |value_omega_0| or any constant multiple of this. Note that magnitude of |omega_0| is not important, the direction it represents is important.
Fischer's Linear Discriminant in Projected Coordinates
=========================
- Claim**
.. |claim1| image:: tex
:alt: tex: \vec{c}=\omega_0={S_W}^{-1}(\vec{m1}-\vec{m2})
.. |prob_matrix1| image:: tex
:alt: tex: \mathbf{Y}\vec{c}=\vec{b}
.. |opt_b_value1| image:: tex
:alt: tex: \vec{b}=(d/d_1, \cdots, <d_1 times>, d/(d-d_1), \cdots, <(d-d_1) times>)^T
|claim1|
is the solution to |prob_matrix1| with |opt_b_value1|
Here is an animation of the 1D example given in class on projections
.. image:: projection1.gif
.. |1DEx1| image:: tex
:alt: tex: \vec{\omega} \bullet y_i + \omega_0 > 0
.. |1DEx2| image:: tex
:alt: tex: \vec{\omega} \bullet y_i + \omega_0 < 0
.. |1DEx3| image:: tex
:alt: tex: \vec{\omega} \bullet y_i < 0
.. |1DEx4| image:: tex
:alt: tex: \vec{\omega} \bullet y_i > 0
Explanation *(feel free to change / correct)*
starts with |1DEx1| for class 1 and |1DEx2| for class 2
the data points are then projected onto an axis at 1 which results in
|1DEx4| for class 1 and |1DEx3| for class 2
one class is then projected onto an axis at -1 which results in
|1DEx4| for all yi
- Support Vector Machines (SVM)**
A support vector for a hyperplane (vector c) with margin (bi >= b) is a sample (yio) such that (*yio = b.) (someone please fix with mathtype)
.. image:: lec11_sv_pic1.bmp .. image:: lec11_sv_pic2.bmp
Support Vector Machines are a two step process: 1) Preprocessing - X1,...,Xd features in kth dimensional real space is mapped to features in n dimensional real space where n>>k. 2) Linear Classifier - separates classes in n dimensional real space via hyperplane.
- Support Vectors - for finding the hyperplane with the biggest margins. - Kernel - to simplify computation (This is key for real world applications)
Previous: [Lecture 10]
Next: [Lecture 12]
