Week 11 Ā· Unit 9.2: My childrenās digital world Ā· 3 lessons
1
What counts as AI?
Digital Literacy 9: Unit 9.2, objectives 9DW.06, 9TC.01
GoalExplain what AI and machine learning are and tell them apart from programs that only follow written rules.
New words
artificial intelligence (AI) Ā· computer systems that do tasks which usually need human thinking, such as recognising speech or picturesmachine learning Ā· building AI by letting a computer find patterns in many examples, instead of writing a rule for every casetraining data Ā· the examples a machine-learning system learns its patterns fromrule-based program Ā· a program that only follows rules that a person wrote in advance
Explanation
Artificial intelligence, or AI, means a computer doing a task that usually needs human thinking, such as understanding speech, recognising a face or sorting messages. A rule-based program follows rules that a person wrote for every case. In machine learning, people give the computer many examples, called training data, and it finds the patterns by itself. A junk-mail filter is a good case. Nobody can write a rule for every trick, so the filter learns from thousands of messages that people marked as junk or not junk. Then it makes a guess about each new message. The guess can be wrong. AI does not understand or feel anything. It matches patterns, and it can sound sure even when it is mistaken.
Kind of program
Who found the pattern?
Example
rule-based
a person wrote every rule
a lamp that switches on at 7 pm
machine learning
the computer, from examples
a filter that learns what junk looks like
Worked examples
Listener
Liked
Suggestion
Nodira
songs A and B
song C?
Tomas
songs A, B and C
ā
A shop app suggests a new song to Nodira. Listeners who liked the same songs as she did also liked it, so the system guessed she would too.
Set-up: save face patternāNew look at the phoneāCompare the patternsāMatch: unlock
Face unlock is a biometric check. At set-up the phone stores a pattern of your face. Each time you look at it, it compares the new pattern with the stored one.
Rashid
Set a timer for fifteen minutes.
Assistant
Timer set for fifty minutes.
Voice assistants and translation tools learned from huge amounts of example speech and text. They can still mishear a word.
Class activity
Rules or examples? (on paper, in pairs). Make twelve small cards with invented message titles such as āFree ticket, click now!ā or āFootball practice moved to 5 pmā. Six are junk, six are normal; keep the answers on the back. Round 1: write at most three rules for spotting junk and test them on all twelve cards. Count the mistakes. Round 2: look at six cards with their answers showing (your training data), find the patterns, then guess the other six, answers hidden. Compare the two rounds. Which tricks fooled your rules? What would happen if your six training cards were all about prizes?
Practice
1
Rule-based or machine learning? (a) a school bell that rings at 9:00 (b) a mail filter that learns from messages people mark as junk (c) a calculator adding two numbers (d) an app that learned to name tree leaves from thousands of photos.
(a) and (c) are rule-based: a person decided exactly what happens. (b) and (d) are machine learning: the computer found patterns in examples.
2
A phone does not unlock for a pupil who has started wearing new glasses. Why might this happen, and does it show the AI is āthinkingā?
At set-up
Today
Result
no glasses
glasses
?
The pattern it sees no longer matches the stored one closely enough. It is comparing patterns, not thinking or knowing who the person is.
3
Why can a programmer not simply write a rule for every junk message?
Senders keep changing their wording and tricks, so there are far too many cases. Learning from many examples copes better, although it can still miss new tricks.
4
A translation tool was trained only on formal books. What might go wrong when a teenager types slang, and why?
Translate: āThat match was sick!ā
It may translate slang wrongly or not at all, because it learned patterns only from formal writing. What the training data contains decides what the system does well.