Web site logical path: [www.psy.gla.ac.uk] [~steve] [educ] [this page]

Hannah Fry's argument

The RI-Xmas lectures 2019 as a model of the multidisciplinarity of important current uses of computation

https://www.bbc.co.uk/iplayer/episodes/b00pmbqq/royal-institution-christmas-lectures

These are aimed at children about 12 years old.

Her lectures (unlike previous series on the whole) did NOT say what the implications were. In fact, unlike most lecturers and against old-fashioned advice, she did NOT tell them what she would say and why; then say it; then tell them what it was that she had just told them.

An interpretation of the implicit argument in her lectures

1. Mathematics: the great thing about maths is the degree of certainty. You can calculate how to do successfully and right first time, things which our ordinary thinking tells us are very unlikely [e.g. sky dive from 30,000 feet without a parachute into a big catch net.] Main problems with this apart from slips in the calculations are if you leave out of your maths model of the situation, a factor which matters.
2. Computers: can do exact calculations much faster than humans. Given time limits in some problems, this can be a critical benefit e.g. matching kidney donors to recipients.
3. Statistics: some important situations do not allow you to calculate deterministically BUT are highly regular when many trials are combined. This is true of a lot of human behaviour. Also true of how infectious disease does and doesn't spread, with or without vaccinations. High certainty collective outcomes from low certainty individual outcomes; and how to exploit this.
4. Probability: correctly combining multiple bits of partial information. Safety in systems by combining faulty sources of knowledge. Autonomous drones; driverless road cars. They typically have multiple sensors, all may fail to detect what they should. These can be combined in two ways: a) AND-ed together: if any one says it's dangerous, shut the vehicle down. b) Bayes theorem says how to combine conditional probabilities; and this can be used to combine sensor outputs to give an accurate total probability given multiple bits of (sources of) partial information. This also lets you use (medical) diagnostic tests that are imperfect, and work out the true chance of you being ill / clear of illness.
5. Machine learning: Programming machines requires you to write down ALL the details of the actions. This is so onerous in many important cases (e.g. how to get a machine to recognise a dog) that it seems more practicable to use AI learning to automate this, giving it only examples => motivation for machine learning. This leads to creating computer programs that were hitherto too hard to write, but also that humans do not know how to write and so do not understand what they are doing.
6. Putting several of the above together can achieve:
   1. [Facebook, Google] taking rather few bits of a person's online behaviour, predict their behaviour and interests with high accuracy.
   2. [used in premier league football] Using a big mass of data on each player's abilities (e.g. running speed, passing accuracy) for each team member AND each of the opposing team's players, it can predict what is the best thing a player can do in any situation (best = most likely to lead to a goal in the next 15 seconds). That is, not just predict someone's voluntary behaviour, but advise which bits of their voluntary behaviour they should select in order to achieve their (team's) objectives / success. (I.e. planning of a whole new kind; not structural and deterministic, but probabilistic.)

Web site logical path: [www.psy.gla.ac.uk] [~steve] [educ] [this page]
[Top of this page]