Machine Learning (part of exam 2/3)

Question 1

Why can it be useful for a machine to learn?

Answer 1

• it’s essential for unknown environments (ie when the designer isn’t omniscent)
• it’s useful as a system construction method (ie expose the agent to reality rather than trying to write down reality)

Question 2

Assign the correct names to the following quotes:

1. Learning is making useful changes in our minds.
2. Learning denotes changes in the system that […] enable the system to do the same task or tasks drawn from the same population more efficiently and more effectively the next time.
3. Learning is constructing or modifying representations of what is being experienced.

Answer 2

1. Minsky
2. Simon
3. Michalski

Question 3

What is Mitchell’s definition of Machine Learning (1997)?

Answer 3

A computer program is said to learn from experience E with respect to some class of task T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.

Question 4

What information needs to be given in order to reach the goal of improving the performance on a task?

Answer 4

• a task T
• a performance measure P
• some experience E

Question 5

On the example of teaching a machine to play Backgammon, what are the task T, the performance measure P and the experience E?

Answer 5

• T: play backgammon
• P: percentage of games won
• E: previously played games

Question 6

What examples have we learning where machine learning is used in our daily life?

Answer 6

• teaching machines to play games
• regognizing spam-Mail
• handwritten character recognition
• classifying stars, galaxies, quasrs,..
• market basket analysis (recommendation systems, store layouts)

Question 7

On the example of Spam-Mail, what are the task T, the performance measure P and the experience E?

Answer 7

• T: sort E-Mails into categories
• P: weighted sum of mistakes (letting spam through is weighted less than misclassifying regular E-Mails as spam)
• E: handsorted E-Mails by user

Question 8

What’s the name of the learning method most spam-filters use to teach their machines to recognize spam mails?

Answer 8

Bayesian Learning

Question 9

On the example of Handwritten Character Recognition, what are the task T, the performance measure P and the experience E?

Answer 9

• T: recognize a handwritten character
• P: recognition rate
• E: MNIST handwritten digit database

Question 10

On the example of Classifying Stars, what are the task T, the performance measure P and the experience E?

Answer 10

• T: classification of celestial bodies
• P: accuracy of classifying
• E: classificatios of astronomers

Question 11

What method is used to classify stars?

Answer 11

learning of multiple decision trees and combining the best rules of each tree

Question 12

On the example of the Market Basket Analysis, what are the task T, the performance measure P and the experience E?

Answer 12

• T: discover items that are frequently bought together
• P: ? possibly revenue of those items
• E: Supermarket check-out data

Question 13

What types of different Learning Scenarios are there?

Answer 13

• Supervised Learning
• Semi-supervised Learning
• Reinforcement Learning
• Unsupervised Learning

Question 14

What is Supervised Learning?

Answer 14

• a lot of labeled examples are provided for training purposes
• machine has to assign labels to examples
• concept learning, classification, regression

Question 15

What is Semi-supervised Learning?

Answer 15

• a few labeled examples are provided for training purposes
• machine has to assign labels to examples

Question 16

What is Reinforcement Learning?

Answer 16

• there are no labeled examples for training purposes
• machine only receives feedback on the labelling assignment it does

Question 17

What is Unsupervised Learning?

Answer 17

• there is no information except the training examples
• clustering, subgroup discovery, association rule discovery

Question 18

Assign the following examples to the correct types of Learning Scenarios.

1. In a video game you find out what to do by how many xp you receive for different actions.
2. You download a few webpages and classify them into various types of webpages. then you tell an algorithm to classify every webpage it finds.
3. An algorithm receives a pack of thousands of tweets and the instruction to sort them into clusters.
4. A handwritten letter is scanned and run through a handwritten character recognition software.

Answer 18

1. Reinforcement Learning
2. Semi-supervised Learning
3. Unsupervised Learning
4. Supervised Learning

Question 19

What is Inductive Learning?

Answer 19

• Given: input x and output f(x) of a function
• Not given: function f
• Problem: given a set of training examples, find a hypothesis h that is as close to the function f as possible, on all examples (so it must generalize from the training examples)
• it ignores prior knowledge
• it assumes that examples are given

Question 20

What is Ockham’s Razor and how does it pertain to curve fitting in the Inductive Learning Method?

Answer 20

“The simplest explanation is often the best/correct explanation.”

When trying to fit a curve to data points, the best curve for machine learning is the curve that is both simple and mostly right (it doesn’t necessarily have to hit all points but should be relatively easy to foresee into the future.)

Question 21

What is Overfitting?

Answer 21

A curve is overfitting if its made to fit all points at the expense of being too complex and inconsistent. The curve can’t realistically be used for other data points because it’s too fitted for the example data points.

Question 22

How can Overfitting be avoided?

Answer 22

Maximize a combination of consistency and simplicity.
Keep a separate validaten set (different from training and test sets) to watch the performance. If the error on the validation set rises, stop training.

Question 23

How does Performance Measurement work? How do we know that we have reached the closest possible solution?

Answer 23

• use theorems of computational and statistical learning theory
• try the solution h on a new set of examples where f is known

Question 24

How did the “Pigeons as Art Experts” experiment (Watanabe et al 1995, 2001) work and what were its findings?

Answer 24

Pigeons were presented with paintings of Chagall and Van Gogh. They received food when they pecked on paintings by Van Gogh.
After some time the pigeons were able to differentiate between the two artists with 95% accuracy when shown paintings they’ve been trained on and 85% accuracy on previously unseen paintings.

Question 25

What are Neural Networks?

Answer 25

They are modelled off the human brain and nervous system. They can extract and detect pattern and generalise them to make predictions.

Question 26

Who were David Hubel and Torsten Wiesel?

Answer 26

They explored the visual cortex’ of cats in the 1960s and discovered line and edge detectors in the visual system.
They received the Nobel prize in 1981.

Question 27

How are biological neurons built and how do they work?

Answer 27

Each neuron/cell has a center (nucleus), a body (soma), dendrites and axons. The dendrites and axons are on the edges of the cell and are responsible for receiving input (dendrites) and sending output (axons).
When a dendrite connects with an axon of another cell, this connection is called a synapse.
If the input reaches a certain strength, it fires, meaning the cell gets activated and passes the activation on to all connected cells along it’s axons.

Question 28

How are artificial neurons built and how do they work?

Answer 28

A neuron is called a node or unit.
The input (in) is weighed: Sum of (all inputs a * all weights w).
The output (a) is determined by the activation function (g).

So we have:
a = g(in)
a = g (sum of (all inputs a * all weights w))

Question 29

What is a perceptron (Rosenblatt 1957, 1960)?

Answer 29

It’s a single node that connects n input signals with one output signal, typically resulting in either -1 or +1.
The activation function is a simple threshold function.

Question 30

How can perceptrons and boolean functions be combined?

Answer 30

The boolean functions “and”, “or” and “not” can be combined by linearly separating the result from the rest.
More comple functions like “xor” can’t be modeled.

Question 31

What is the Perceptron Learning Rule for Supervised Learning?

Answer 31

It’s a function that includes a learning rate (alpha) and an error calculation, which calculates the weights needed for given inputs and outputs.

Question 32

How can the error of a network be measured?

Answer 32

The error of one training example x can be measured by the squared difference of the output value h(x) and the desired target value f(x).

For evaluating the performance of a network we can try the network on a set of datapoints and average the values (values are calculated like above).

Question 33

How does a machine find the correct weights for given inputs and outputs?

Answer 33

It’s almost always a search problem. The machine has to search for the right values and it can use different methods for that, e.g. heuristic search functions, local search functions,..

Question 34

What is Hypothesis Space?

Answer 34

It’s the space of all possible values for the weights we’re looking for.

Question 35

Out of all the values for weights in the hypothesis space, how do we know which ones are better than others?

Answer 35

We use an evaluation function that measures the error of each weight. So we search for weights that have a low error on the training data.

Question 36

What is the best weight setting for one example in an error landscape?

Answer 36

It’s where the error measure for this example is minimal.

Question 37

How can we find the best weight setting in an error landscape?

Answer 37

via Gradient Descent: go downhill in the direction where it is steepest (hill-climbing search).

Question 38

Why is the regular threshold activation function not useable in machine learning?

Answer 38

Because it’s not differentiable.

(??? honestly i don’t get this at all but this is what he said in the lecture and what it says on the slides so ¯_(ツ)_/¯ there is a lot of math involved here. If you wanna check it out for yourselves, it’s in the video of the lecture of 13.11.2024 at approx 1:18:00. have fun!)

Question 39

What is the commonly used activation function and why?

Answer 39

The Sigmoid Activation Function. Because it’s easy to differentiate and non-linear.

Question 40

True or False: Every continuous function can be modeled with three layers (incl one hidden layer).

Answer 40

True

Question 41

What is Backpropagation Learning?

Answer 41

Usually we compute the error of a function but if there is a hidden layer, we need to calculate the error of the output layers, then backpropagate that to the hidden layer.
Delta is the error term of the output node times the derivation of its inputs.
This is done to update the weights and minimize errors, so in a way, this is how a machine learns.

Question 42

What is Deep Learning?

Answer 42

It’s a neural networks with many hidden layers.

Question 43

In what area has Deep Learning made great successes and how?

Answer 43

Image classification.
The many layers are fully connected (every node on one level is connected to every node on the next level). There are layers specifially trained on recognizing i.e. edges, corners, diagonal lines, faces, trees,…

Question 44

What is needed for Deep Learning?

Answer 44

• a lot of training data (big data)
• fast processing
• unsupervised pre-training of layers

Question 45

What is Convolution in Neural Networks?

Answer 45

It’s a technique in image processing, where for each pixel of an image a new feature is computed using a weighted combination of its neighborhood (n*n pixels around it).
Depending on the weights this could blur the image or have it show only the edges, etc.

Question 46

What is Neural Artistic Art Transfer?

Answer 46

Using Deep Learning an input image can be altered to look like a reference image (ie look like as if Van Gogh had painted it).

Question 47

What are Generative Adversarial Networks (GANs)?

Answer 47

They’re methods for increasing robustness of a learning machine. “Invisible” changes are made to an image to confuse the machine, to make mistakes and further down the road, to make the machine recognize such mistakes.

Question 48

What are Recurrent Neural Networks?

Answer 48

They allow to process sequential data by feeding back the output of the network into the next input.