Lecture 1 – 31/01 – Introducti on
Topics
- What is machine learning? How to evaluate (basis level)?
- Social implication (bias, adoption, privacy/trust, skills)
- Hype and technology adoption
- Panic and resistance
- A deep dive into the influence of technology
Lecture & seminar content
Will:
- Lectures: machine learning, technology acceptance, digital inequality
- Seminars: stakeholder analysis, assessing the quality of existing systems
Lennert:
- Lectures: media effects, technology influence
- Seminars: discussion, gaining a deeper understanding of the theory
This course
Combination of language (linguistics, anthropology), society (communication, sociology) and
technology (computer science, engineering).
Hypes and panics
A hype is when people are overly excited about some new technology.
Examples are self-driving cars or Artificial Intelligence. A panic is when
people are overly worried about some new technology, such as filter
bubbles. These two phenomenon might fomm a feedback loop.
Questions for these course:
- How can we find the right balance between hype and panic?
- How can we assess the true effects and capabilities of smart technology?
For lack of a better term: technology realism.
Examples of language technology
1. Chatbots
2. Voice assistants
3. Search engines
, Lecture 2 – 07/02 – How does ‘smart’ technology work?
Outline
- Modularity
- Algorithms
- Machine learning
- Evaluation metrics
- Natural language processing
Terminology
Artificial intelligence: using computers to solve problems
previously thought to be only possible through the human
brain (pattern detection, learning, language production).
Machine learning: the study of computer algorithms that
allow computer programs to automatically improve through
experience.
Natural language processing: the application of
computational techniques to the analysis and synthesis of
natural language and speech.
How does smart technology work?
Example: Google Nest. It is a modular system with many different components, where each
component is responsible for a sub-task:
- Recognize speech (the actual words)
- Understand language (the whole meaning)
- Pre-defined procedures: setting the alarm, playing a song
- Generate language (meaning and syntax)
- Produce speech (speech synthesis)
These tasks work via algorithms: a step by step procedure for solving a problem or
accomplishing some end.
Algorithm properties
Traditionally, algorithms are:
- Fully pre-specified procedures: all the rules are written beforehand
- With predictable behavior: you know the rules, you know the behavior
, - Because of that: (in principle) verifiable: you can double check the results because you
know how the algorithm works.
- Because of that: reliable, in predictable environments: you can be sure they work the
same way each time.
For a certain range of inputs, we will obtain specific kinds of results.
Rule-based systems are great, but …
1. You need to specify exactly what the computer should do. E.g. how to recognize the
picture of cats? Very difficult to capture in rules.
2. You need to anticipate what kinds of input users will provide. E.g. what will people
want to talk about? Doable in theory (especially in restricted contexts), but very time-
consuming.
Machine learning
It is the study of computer algorithms that allow computer programs to automatically improve
through experience.
Statistcal modelling Machine learning
Limited amount of data Lots of data
Variables Features
Explanation Prediction
Theory-driven Data-driven
Focus on interpretation Focus on accuracy
Types of machine learning
There are roughly 3 types of machine learning:
- Supervised learning: learning from tagged data (example pairs)
- Unsupervised learning: learning from untagged data
- Reinforcement learning: learning to take actions to maximize some reward.
, In this course the focus lies on supervised learning. Though many advanced systems today
learn on untagged data (they are unsupervised) or use a combination of techniques
(ChatGPT).
Classification
Supervised: labelled data. The outcome is binary or categorical: yes/no, cat/dog/hamster.
Examples: spam filter of classification of images of fruit (pear, apple, banana)
Regression
Supervised: labelled data. The outcome is continuous (numerical) examples: prediction of
essay score, house prices, car sales in a year.
Types of machine learning – practice
1. Is the prediction binary/categorical -> classification
2. Is the prediction continuous -> regression
3. Does the prediction seek to group similar things -> clustering
Classification, regression, or clustering
1. Cancer cell detection -> classification
2. Social network analysis -> clustering
3. Face detection -> classification
4. Predicting number of applications to our MA program -> regression
5. Group of shopping items -> clustering
Machine learning pipeline – supervised learning
