Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                
Scribd Logo
Upload

Welcome to Scribd!

  • 1 views

    4.11 HW Machine Learning SOLUTION

    Uploaded by

    ahihi channel

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    4.11 HW Machine Learning SOLUTION

    Uploaded by

    ahihi channel
    1 views2 pages

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    1 views2 pages

    4.11 HW Machine Learning SOLUTION

    Uploaded by

    ahihi channel

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 2

    4.

    11 HW Machine Learning SOLUTIONS


    Q1: Discuss and demonstrate in detail how an accurate classifier can be found for the following training data
    shown in 2D.
    Many possible kernel transformations can be
    done, this one below is from the LMS x y z=x+y2 shape
    4 2 8 triangle
    1 3 10 triangle
    3 3 12 triangle
    6 3 15 triangle
    8 4 24 triangle
    2 5 27 triangle
    4 5 29 circle
    6 5 31 circle
    4 6 40 circle
    6 7 55 circle

    z=(x-5)2+(y-6)2 will also work as a kernel


    transformation for this dataset x y z=(x-5)2+(y-6)2 shape
    where z > 6 forms a hyperplane separator which 4 2 17 triangle
    1 3 25 triangle
    is a plane.
    3 3 13 triangle
    6 3 10 triangle
    8 4 13 triangle
    2 5 10 triangle
    4 5 2 circle
    6 5 2 circle
    4 6 1 circle
    6 7 2 circle

    The data is not linearly separable so a perceptron is not suitable for classification in this case. An SVM can be used
    and a transformation function (kernel) can be applied to the data for example z=x+y2 then a linear classifier can be
    applied in the transformed space, this will result in a decision boundary hyperplane in 3D for example z > 28 are
    circles. (NOTE: Other valid kernel transformations may exist)

    Q2: The Abacus company came up with the following hyperplane as shown in the diagram below. What are the
    issues with the result?
    If the training data was sufficient and representational then underfitting has
    occurred on the training data as the hyperplane boundary is not classifying
    correctly and it is unlikely that the predictions on new data will be correct.
    There may not have been enough training data, or it may not have been
    representational of the population to determine, there may be outliers that
    are misclassified.

    Consider the following table for an automated seat belt fastener alert in a car which is making decisions based on a
    function in a single hidden neuron that accepts weighted input from two binary sensor inputs and a conditional
    function applied to them.
    Input1 =(weight sensor on seat > 1kg) Input2= (seatbelt on for seat) Show Alert =
    X1 X2 NOT(NOT(Input1) OR (Input2))
    0 0 0
    0 1 0
    1 0 1
    1 1 0

    CHES Algorithmics Unit 4


    4.11 HW Machine Learning SOLUTIONS

    Q3: Draw a clearly labelled neural network and its structures that represents the automated seat belt fastener
    alert decision.

    Show input, hidden and output layers

    The inputs and their alert outcomes for the seat belt alert mechanism can also be shown on a cartesian plane below.

    Q4: Can this seat belt alert be represented by a perceptron neural network and/or a support vector machine or
    both? Explain your answer.
    • The inputs and their results are linearly separable when plotted on a cartesian plane so can be decided using a
    support vector machine. The SVM classifier would be a hyperplane that separates the Nos and Yesses with the
    biggest margin possible.
    • The inputs and their results are linearly separable as the output of the function in the hidden layer is
    true/false and can also be represented by a perceptron neural network.
    • Since the only data points possible for this scenario are Boolean values (0,1), (0,0), (1,1) and (1,0) a
    deterministic algorithm can easily determine the seat belt alert.
    If (NOT(NOT(Input1) OR (Input2))) then
    Show alert
    End if

    Q5: If you answered yes to any of the previous then show the classifiers for each Machine Learning method(s).
    • Perceptron Neural Network Binary classify O=( true if w1x1 + w2x2 > 0, false if w1x1 + w2x2=0)
    • SVM hyperplane to classify the two types of output with the biggest margin possible would go through the
    cartesian points (0.5,0) and (1,0.5)
    Hyperplane gradient=1 w1(0.5)1-w1(0)-b=0
    Weight vector gradient = -1 w1(0)-w1(0)-b=-1  b=1
    hence w1=-w2 -w1=w2 w1(1)-w1(0)-b=+1  w1=2
    w1x1+w2x2-b=0 on hyperplane w2=-2
    w1x1+w2x2-b=-1 for (0,0),(1,1)
    w1x1+w2x2-b=1 for (1,0) 2x1-2x2-1=0 defines
    hyperplane for SVM

    CHES Algorithmics Unit 4

    You might also like