The Machine Learning Landscape

Question 1

What is a definition for “Machine Learning”?

Answer 1

Machine Learning is the field of study that gives computers the ability to learn how to solve a specific problem, without being explicitly programmed.

Question 2

What is the definition of “Training Set” and “Test Set”?

Answer 2

• The Training Set is the examples/instances that the system uses to learn from.
• The Test Set is the instances that you uses to test the system on what he learnt from the training set.

Question 3

What is the definition of a “Model”?

Answer 3

The part of the machine learning system that learns and makes predictions.
Examples: Neural Networks, Random Forest.

Question 4

In which cases Machine Learning is a great solution?

Answer 4

1. Problems for which existing solutions require of lot of fine-tuning or complex rules;
2. Complex problems for which using a traditional approach yields no good solution;
3. Fluctuating environments (a ML system can easily be re-trained on new data);
4. Getting insights about complex problems and large amounts of data.

Question 5

What are the main criteria used to classify Machine Learning systems?

Answer 5

1. Training Supervision (supervised, unsupervised, semi-supervised, self-supervised…);
2. Incremental/Online or Batch Learning;
3. Learning Approach (instance-based or model-based).

Question 6

What is the definition of “Supervised Learning”?

Answer 6

The model is trained using instances with their features (called predictors or attributes) and the desired solutions (called labels).
Examples: classification, regression: predict a target numeric value.

Question 7

What is the definition of “Unsupervised Learning”?

Answer 7

The model is trained without any solutions (unlabeled).
Examples: clustering, anomaly detection, association rule learning, dimensionality reduction.

Question 8

What is the definition of “Clustering”?

Answer 8

The task of grouping a set of objects in such a way that objects in the same group (called cluster) are more similar (in some way) to each other than to those in other groups (clusters).

Question 9

What is the definition of “Dimensionality Reduction”?

Answer 9

The task of reducing the number of features in a dataset while retaining as much information as possible. It is a process of transforming high-dimensional data into a lower-dimensional space that still preserves the essence of the original data.

Question 10

What is the definition of “Anomaly Detection”?

Answer 10

The task of identifying rare instances or observations which can raise suspicions by being statistically different from the rest of the observations.
Examples: credit card fraud, manufacturing defaults.

Question 11

What is the definition of “Self-Supervised Learning”?

Answer 11

The model trains itself to learn one part of the input from another part of the input: it generates a fully labeled dataset from a fully unlabeled one.

Question 12

What is the definition of “Reinforcement Learning”?

Answer 12

The learning system, called an agent, can observe the environment, select and perform actions, and get rewards in return (or penalties as negative rewards). It must then learn by itself what is the best strategy, called a policy, to get the most reward over time.

Question 13

What is the definition of “Batch/Offline Learning”?

Answer 13

The system is incapable of learning incrementally: it must be trained using all the available data. First the system is trained, and then it is launched into production and runs without learning anymore; it just applies what it has learned.

Question 14

What is the definition of “Association Rule Learning”?

Answer 14

The task of detecting dependency of one data item on another data item and maps accordingly so that it can be more profitable. It tries to find some interesting relations or associations among the variables of dataset.

Question 15

What is the definition of “Semi-Supervised Learning”?

Answer 15

The model is train with partially labeled instances. Most of these models are combinations of unsupervised and supervised algorithms. This type of learning is interesting as labeling is usually time-consuming and costly.

Question 16

What is the definition of “Model Rot” or “Data Drift”?

Answer 16

The phenomenon by which a offline model’s performance tends to decay slowly over time because the world continues to evolve while the model remains unchanged.

Question 17

What is the definition of “Incremental/Online Learning”?

Answer 17

The system is trained incrementally by feeding it data instances sequentially, either individually or in small groups called mini-batches.

Question 18

What is the “Learning Rate”? And what is the difference between a high and low one?

Answer 18

The speed at which the online learning model adapt to new and changing data:
- High: rapidly adapt to new data BUT tends to quickly forget the old data;
- Low: learn more slowly BUT less sensitive to noise and outliers.

Question 19

What is the definition of “Instance-based Learning”?

Answer 19

One of the 2 main approaches to generalization (how to react in front of new instances).

The system learns the examples by heart, then generalizes to new cases by using a similarity measure to compare them to the learned examples (or a subset of them).

Question 20

What is the definition of “Model-based Learning”?

Answer 20

One of the 2 main approaches to generalization (how to react in front of new instances).

The system builds a model (like a linear regression) through the examples and then use that model to make predictions.

Question 21

What is the “Performance Measure”?

Answer 21

The metric used to determine how good or bad the model is performing.

Question 22

What is the difference between a “Utility Function” and a “Cost Function”?

Answer 22

A “Utility Function” determines how GOOD the model is while a “Cost Function” determines how BAD it is.

Question 23

What type of algorithm would you use to allow a robot to walk in various
unknown terrains?

Answer 23

The best Machine Learning algorithm to allow a robot to walk in unknown terrain is Reinforced Learning, where the robot can learn from response of the terrain to optimize itself.

Question 24

What type of algorithm would you use to segment your customers into multiple
groups?

Answer 24

The best algorithm to segment customers into multiple groups is either supervised learning (if the groups have known labels) or unsupervised learning (if there are no group labels).

Question 25

Would you frame the problem of spam detection as a supervised learning problem or an unsupervised learning problem?

Answer 25

Spam detection is a supervised learning problem because the labels are known (spam or no spam).

Question 26

What is “Transfer Learning”?

Answer 26

The process of transferring knowledge learned from one task to another in order to boost its performance.
Examples: if the model knows how to differentiate animal species through pictures, the next step would be to classify them.

Question 27

What is a “Policy” in reinforcement learning?

Answer 27

The strategy or action the agent should choose when it is in a given situation in order to get the most reward.

Question 28

What is the main risk with “Batch Learning”?

Answer 28

The main risk is “Model Rot” or “Data Drift”.

Question 29

When is it difficult of impossible to use a “Batch Learning” model?

Answer 29

1. The amount of data is huge;
2. The data is fast-evolving.

Question 30

When is it difficult of impossible to use a “Batch Learning” model?

Answer 30

1. The amount of data is huge;
2. The data is fast-evolving;
3. The computing resources are limited.

Question 31

What is the definition of “Out-of-Core Learning”?

Answer 31

A system that can handle data that cannot fit into your computer memory. It uses online learning system to feed data in small bits.

Question 32

When is it difficult of impossible to use a “Batch Learning” model?

Answer 32

1. The amount of data is huge;
2. The data is fast-evolving;
3. The computing resources are limited.

Question 32

What type of learning algorithm relies on a similarity measure to make predictions?

Answer 32

Learning algorithm that relies on a similarity measure to make predictions is instance-based algorithm.

Question 33

What are the main “bad data” issue categories?

Answer 33

1. Insufficient quantity of training data
2. Non-representative training data
3. Poor-quality data (errors, noise, outliers)
4. Irrelevant features/attributes

Question 34

What are the main “bad model” issue categories?

Answer 34

1. Overfitting the training data
2. Underfitting the training data

Question 35

What is “Sampling Bias”?

Answer 35

A bias where certain groups of individuals are more likely to be included in a sample than others, leading to an unrepresentative sample.

Question 36

How to avoid using irrelevant features in a model?

Answer 36

Through using “Feature Engineering”:
1. Feature selection (selecting the most useful features among existing ones);
2. Feature extraction (combining existing features to produce a more useful one);
3. Creating new features by gathering new data.

Question 37

What is “Overfitting” and when does it happen?

Answer 37

It means that the model performs well on the training data, but it does not generalize
well. Overfitting happens when the model is too complex relative to the amount and noisiness of the training data

Question 38

How to avoid overfitting?

Answer 38

1. Simplify the model by:
   
   • selecting one with fewer parameters,
   • reducing the number of attributes in the training data,
   • by constraining the model;
2. Gather more training data;
3. Reduce the noise in the training data (e.g., fix data errors and remove outliers).

Question 39

What is “Regularization”?

Answer 39

Constraining a model to make it simpler and reduce the risk of overfitting

Question 40

What is a “Hyper-Parameter”?

Answer 40

The amount of regularization to apply during learning can be controlled by a hyper‐
parameter. A hyperparameter is a parameter of a learning algorithm (not of the
model). As such, it is not affected by the learning algorithm itself; it must be set
prior to training and remains constant during training.

Question 41

What is “Underfitting” and when does it happen?

Answer 41

It means that the model is too simple to learn the underlying structure of the data. For example, a linear model of life satisfaction is prone to underfit; reality is just more complex
than the model, so its predictions are bound to be inaccurate, even on the training
examples.

Question 42

How to avoid underfitting?

Answer 42

1. Select a more powerful model, with more parameters;
2. Feed better features to the learning algorithm (feature engineering);
3. Reduce the constraints on the model (for example by reducing the regularization
   hyperparameter).

Question 43

What do model-based learning algorithms search for? What is the most common strategy they use to succeed? How do they make predictions?

Answer 43

Model based learning algorithm search for the optimal value of parameters in a model that will give the best results for the new instances. We often use a cost function or similar to determine what the parameter value has to be in order to minimize the function. The model makes prediction by using the value of the new instance and the parameters in its function.

Question 44

If your model performs great on the training data but generalizes poorly to new instances, what is happening?

Answer 44

If the model performs poorly to new instances, then it has overfit on the training data.