Zoom Logo

Bristol & Bath Seminar Series - Professor Gerd Gigerenzer - Shared screen with speaker view
Paul Bain
22:27
A second-order question...who is in a better position to judge whether a "real-world" situation is stable or uncertain? A human or a computer?
Paul Bain
39:33
But the flu recency heuristic doesn’t seem that good for planning (e.g., resource procurement and distribution, staffing), which is why you need to make predictions. You would need more than a week in advance…
Edwin Dalmaijer
40:44
@Paul: I think GFT uses current-week search data, so it wouldn't add a whole lot to long-term planning. Also, you could use the heuristic current year == last year for longer-term prediction?
Ed(he/him) - PG Psychology
42:06
@Edwin, I think you're right, the law of recency doesn't require the previous week, just most recent data. So if you wanted annual predictions, you could use the previous year?
Paul Bain
43:12
@Edwin. Good to know. The longer-term heuristic "could" still be better, but I wonder what the effect on the error is for heuristics v more complex models for the longer timeframe that matches planning timeframes needed. I don't see why it would necessarily be better.
Marin Dujmovic
43:42
Wouldn't that lead to similar problems as the Google approach. If the swine flu happened last year but doesn't affect this year then you overshoot. If the swine flu did not happen last year but happens this year you underestimate.
Marin Dujmovic
44:37
Obviously, if they're similar in performance - the simpler model is preferable anyway.
Lucas Muller
45:12
You could simply adjust the recency heuristic to need. E.g. on a 3-month basis or a 6-month basis. The longer the time frame you use as a reference, the smaller the influence from one-off events like the swine flu.
Paul Bain
59:26
I don't get the heuristic advantage for the turkey illusion. Based on the most recent day/week approach, you make a tragically wrong prediction. Based on a more complex model (understanding feasts and other contextual info) you would make a better prediction...
Marin Dujmovic
01:01:10
Recency is only one heuristic. For the Turkey it could be something else. E.g. I as a newcomer look at the others besides me and see that there is not a single Turkey older than 6 months - so something is probably going to change - and I know it without too much complex data.
Paul Bain
01:01:52
GOod point...I also started to think I missed something in the presentation of that point...
timle
01:01:57
maybe referring to the evolutionary aspect of caution and fear being subverted by a bad predictive model
Benjamin Woolf
01:14:26
Could I ask a follow up? If you choose the heuristics based of some type of empirical methods would you not then risk some type of overfitting issues (assuming you are choosing from many heuristics)? But if you are worried about Knightean uncertainty, woudln't an analytic method of choosing heuristics fail - because the uncertainty by definition can't be quantified and included in a mathematical model?
Roland Baddeley
01:18:13
https://openai.com/blog/deep-double-descent/
Roland Baddeley
01:19:09
I think the above is one of the more important results in the whole of machine learning!
Edwin Dalmaijer
01:19:45
Great talk!
James Elson
01:19:49
Thank you :)
Paul Bain
01:19:52
thank you!
Benjamin Woolf
01:19:53
thank you -fascinating talk
Lucas Muller
01:19:54
Thank you!
Jeffrey Bowers
01:19:55
thanks so much!
Aoife Bailey
01:19:56
Thank you!
Dan Beechey
01:19:59
Thank you!
Angela Rowe
01:20:00
thanks
Roland Baddeley
01:20:01
Tanks you
Tom Cannon
01:20:02
thanks Gerd, appreciated
Chris Jarrold
01:20:03
Thanks very much - a wonderful talk
Milton Montero
01:20:05
Thanks!!
catherinenaughtie
01:20:05
Thank you
Sophia Jones
01:20:06
Thank you! Very interesting talk
hani akasheh
01:20:07
thanks
Rumeysa Kuruoglu
01:20:08
thank you
Jon Uphoff
01:20:09
Thank you Gerd!
Ed(he/him) - PG Psychology
01:20:09
Thank you Dr Gigerenzer!
timle
01:20:14
thank you
Benjamin Evans
01:20:19
Thanks!