A short quiz on machine learning in computer vision

Adrian F. Clark

1. Why might one "polish" the result of a genetic algorithm?

   • because the GA gives its result in the wrong coordinate system

   • to improve the accuracy of the solution

   • because the GA gives the wrong result
   • because a numerical optimizer is too inaccurate

   A GA is excellent at exploring the breadth of a search space and finding an approximate solution but it is less effective at determining the exact solution. Using the GA's result to initialize a numerical optimizer, which is what "polishing" means here, combines the best of both algorithms.

2. What is over-training in neural networks?

   • where the number of training epochs is too large
   • where the network recognizes only general patterns

   • where the network recognizes specific patterns and does not generalize

   • where the network fails to converge

   Over-training is when the neural network 'learns' each input pattern but is unable to classify correctly similar patterns: it has 'forgotten' the general pattern by focussing on the specific training ones.

3. Which of the following is NOT a well-known deep neural network architecture?

   • ResNet
   • AlexNet
   • GoogLeNet

   • DNNarch

   AlexNet was the first deep CNN to show real promise, and this was followed by GoogLeNet. ResNet is roughly the state of the art, and RecNet is one I make up!

4. What is the general strategy of a GA?

   • the best chromosomes 'reproduce' randomly

   • evolution proceeds by mutation alone
   • all chromosomes 'reproduce' randomly
   • the worst chromosomes 'reproduce' randomly

   In a GA, the best chromosomes (those with the lowest costs) 'reproduce' to produce the next generation. It is common to mutate a small number of these 'children' randomly, and to copy some of the best parents from one generation to another. Some forms of GA do permit only mutation by these converge on a solution much more slowly than others.

5. Which of the following is an unsupervised learning technique?

   • k-nn

   • k-means

   • WISARD
   • CNN

   k-means is unsupervised, while both WISARD and CNN are supervised learning techniques. k-NN isn't really a machine learning technique, more a way of matching things for similarity.

6. How goes a real-valued genetic algorithm work?

   • it adds together randomly-chosen parameters

   • it interpolates along lines in parameter space by a random amount

   • it multiplies together randomly-chosen parameters
   • it repeatedly sets parameters to random values

   Real-valued GAs work by choosing two 'genes' randomly to give two points in parameter space, then interpolating along the line joining them by a random amount.

7. How does a GA measure effectiveness on a problem?

   • it uses a cost function to measure the effectiveness of one chromosome
   • it uses crossover to measure effectiveness
   • it uses mutation to measure effectiveness

   • it uses a cost function to measure the effectiveness of all chromosome

   A result is computed for each chromosome (set of values) and the ones closest to the correct solution -- the one with the lowest cost -- are better.

8. What is the purpose of the RELU layer in a CNN?

   • it down-samples regions of its input
   • it determines the final result
   • it applies the convolution coefficients to an image

   • it applies the activation function

   The RELU layer applies the activation function to each element independently.

9. Which of the following is a deterministic machine learning technique?

   • MLP
   • GP
   • WISARD

   • SVM

   All the techniques listed except SVM are stochasic, meaning they involve the use of random numbers and hence normally produce different results every time they are run on the same data. SVM doesn't do this and so is deterministic, producing the same output every time the same input is used.

10. How does SVM work?

    • it reduces the problem to a linear one and finds the best straight line

    • it finds the hyperplane that maximizes the distance between classes

    • it finds the smallest hypersphere that minimises encloses each class
    • it finds the hyperplane that minimizes the distance between classes

    SVM maximizes the distance between classes and uses a hyperplane to separate them.