Nothing Special   »   [go: up one dir, main page]
Scribd Logo
Upload

Welcome to Scribd!

  • 52 views

    Ell784 Aq

    Uploaded by

    lovlesh roy
    This document contains a multiple choice quiz for a machine learning course. It has 4 multiple choice questions about error functions, feature selection, dimensionality reduction techniques, and regularization. It also has 2 numerical questions about evaluating clustering accuracy from a confusion matrix and interpreting results from training an SVM on different parameter settings. The questions assess understanding of core machine learning concepts covered in the course.

    Copyright:

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

    Ell784 Aq

    Uploaded by

    lovlesh roy
    52 views2 pages
    This document contains a multiple choice quiz for a machine learning course. It has 4 multiple choice questions about error functions, feature selection, dimensionality reduction techniques, and regularization. It also has 2 numerical questions about evaluating clustering accuracy from a confusion matrix and interpreting results from training an SVM on different parameter settings. The questions assess understanding of core machine learning concepts covered in the course.

    Original Title

    ell784_aq

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document contains a multiple choice quiz for a machine learning course. It has 4 multiple choice questions about error functions, feature selection, dimensionality reduction techniques, and regularization. It also has 2 numerical questions about evaluating clustering accuracy from a confusion matrix and interpreting results from training an SVM on different parameter settings. The questions assess understanding of core machine learning concepts covered in the course.

    Copyright:

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

    Ell784 Aq

    Uploaded by

    lovlesh roy
    This document contains a multiple choice quiz for a machine learning course. It has 4 multiple choice questions about error functions, feature selection, dimensionality reduction techniques, and regularization. It also has 2 numerical questions about evaluating clustering accuracy from a confusion matrix and interpreting results from training an SVM on different parameter settings. The questions assess understanding of core machine learning concepts covered in the course.

    Copyright:

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

    ELL784/EEL709: Introduction to Name:

    Machine Learning
    Assignments Quiz Student Number:
    Please answer on this question paper itself; TA:
    do not use a separate answer book
    Date:
    Maximum marks: 12, Form: A

    Section 1. Multiple choice questions

    Each question may have any number of correct choices, including zero. Encircle all choices you believe
    to be correct (1 mark for each correct choice, -0.5 for each incorrect choice). No justification is required.
    1. We have seen two error functions that are used for neural networks: sum-of-squares error (SSE), and
    cross-entropy error (CEE). Suppose we are training a neural network for binary classification. Which
    of the follwing are true?
    (a) SSE cannot be used; it works only for regression.
    (b) CEE should be preferred to SSE, because CEE is closer to classification error, which is what
    we really care about.
    (c) CEE should be preferred to SSE, because CEE also takes into account the magnitude of error,
    rather than just right/wrong.
    (d) Both CEE and SSE can give good results, but in principle CEE might be slightly preferable
    because it corresponds to maximising the likelihood of the data.
    2. When doing multi-variable regression (multiple inputs, single output), which of the following might
    be reasonable ways of doing feature selection, assuming you are confident the features are largely
    independent of each other?
    (a) Try eliminating features one-at-a-time; pick the features that lead to the greatest increase in
    training error upon elimination.
    (b) Normalise each feature to have mean 0 and variance 1; then just pick the features with the
    highest positive weights in the full model.
    (c) Try eliminating features one-at-a-time; pick the features that lead to the greatest increase in
    testing error upon elimination.
    (d) Pick the features with the highest absolute weights in the full model trained on the raw inputs.
    (e) Try eliminating features one-at-a-time; pick the features that lead to the greatest increase in
    cross-validation error upon elimination.
    3. You have seen how both a neural network and PCA can be used to map data to a lower-dimensional
    representation. Suppose, for your labeled data set, you contstruct two representations: one consisting
    of the top M principal components, and one via training a neural network with a single hidden layer
    with M units. Now, you use both these M -dimensional representations as input features to train an
    SVM to classify your data. Which of the following are true?
    (a) Generally, the neural network representation should work better.
    (b) Generally, the PCA representation should work better.
    (c) There is no general or systematic reason why one representation should work better than the
    other.
    (d) The parameter settings for the SVM will generally determine which one works better.

    1
    4. You are doing regularised polynomial regression, and you find that your current model has similarly
    high error during both training and testing. Which of the following might generally be expected to
    reduce error?
    (a) Since this is underfitting, I can try decreasing the number of training data points to get a
    better fit.
    (b) Since this is overfitting, I can try increasing the number of training data points to get a better
    fit.
    (c) Increasing the order of the polynomial.
    (d) Using less noisy training data.
    (e) Weakening the regularisation.

    Section 2. Numerical questions

    5. Consider the following confusion matrix for a three-class data set on which an unsupervised clustering
    has been done:

    Actual class 1 Actual class 2 Actual class 3


    Placed in cluster 1 31 34 2
    Placed in cluster 2 19 3 34
    Placed in cluster 3 1 19 17
    Placed in cluster 4 9 43 12

    As you did in the assignment, evaluate the overall accuracy of this clustering by identifying each cluster
    with the most frequent label in it. [2]

    6. Consider the following results from a soft-margin RBF SVM with parameters C and σ:
    C σ Training error Cross-validation error C σ Training error Cross-validation error
    1 5 0.25 0.26 10 5 0.17 0.24
    1 10 0.28 0.29 10 10 0.21 0.21
    5 5 0.21 0.22 50 5 0.06 0.37
    5 10 0.23 0.24 50 10 0.12 0.32
    (a) Give the values of C and σ for the best-fit case. [1]

    (b) In the above table, how many cases each of overfitting and underfitting are there, relative to the
    best-fit case? [1]

    (c) Suppose you are at an underfit case. Should you increase or decrease C, in order to try and improve
    the fit? [1]

    (d) In terms of the bias-variance trade-off, which of the given parameter settings corresponds to the
    highest bias? [1]

    You might also like