Introduction to Machine Learning (COMP30027 W1 L1)

How many buildings? How many pedestrians? -> Too much to answer in short time

Problem

• Lots of data
• How to use it?
  • Which parts are meaningful & what parts are noise?
  • How do you extract useful info from the data?

Definition of machine learning

Automatic extraction of valid, novel, useful, and comprehensible knowledge (rules, regularities, patterns, constraints, models, …) from arbitrary sets of data

ML tasks

• Classification
• Clustering
• Regression
• Probability estimation
• Sequence discovery
• Association rule mining
• Model fitting
• …

ML vs. data mining / data science?

• Machine learning tends to:
  • Focus on theory more than application (Why do we apply this model to this task?)
  • Ignore problems of run time/complexity unlike data mining
• Data mining tends to:
  • Focus more on applications than theory
  • Worry about run time and scalability
• Data science tends to:
  • Combine elements of both
  • Focus on interpreting and communicating data insights

… but there’s a lot of overlap between all three

Tutorial: theoretical concepts and methods cover (work through numerical examples, understand how and why the algorithms work. Similar to the final exam)

Practical: hands-on experience applying machine learning methods (scikit-learn, experiment and see results on small data sets. Helpful for assignments)

What is learning?

• “Fitting a function to data: finding the relationship between input and output”
• output = f(input)
• Why do we learn?
  • Learning makes it possible to generalise: produce an output for any input, even inputs you’ve never seen before

Why generalise?

• When don’t you need to learn?
  • If the input set is finite (one to one), then you can memorize all the mappings from input->output. (small input set)
  • When you cannot derive anything from the input&output. If there is no rule that could relate input to output.
  • However, memorization does not work for real-world problems.
    • Because
      1. input is continuous (infinite values) and
      2. we want to predict future events

Machine learning tasks

• Classification – predict discrete class labels based on features
• Regression – predict continuous outcomes based on features
• Predict relationships between features / outcomes (sequence learning, association rule mining)
• Understand and reconstruct the processes that produced the features / outcomes (model fitting, probability estimation)

Specific machine learning methods

• Naïve Bayes classifiers
• Decision trees
• Support vector machines
• Linear regression
• Logistic regression
• Gaussian mixture models
• Hidden Markov models
• Perceptron
• Deep neural networks