3 ms·
Can someone TL;DR what superficial cues are in this paper?
by freemint 4y ago
Can someone TL;DR what superficial cues are in this paper?
- DiabloD3 4y agohttps://www.gwern.net/Tanks https://www.gwern.net/Tanks
- Hugsun 4y agogwern is the opposite of TL;DR
- PaulHoule 4y agoIt could be anything that's predictive or seems predictive of the outcome because of bias, accidents in the environment, etc. See https://en.wikipedia.org/wiki/Clever_Hans https://en.wikipedia.org/wiki/Clever_Hans
- freemint 4y agoIn this paper not somewhere else. Ok, more precisely In the datasets used for text comprehension
- PaulHoule 4y agoI've worked with this kind of dataset and I'd say that the relationship between cues and real understanding is like that between spam and real emails in many people's inboxes. To take the example of Clever Hans, cues are just ubiquitous everywhere and a system that learns to see them really well is going to conclude "Tyrone is a thug" because heuristics are easy and justice as hard. It deeply problematizes the field of machine learning based text "understanding" because it is impossible to remove the cues just as Clever Hans problematizes work on animal intelligence. Looking for "understanding" in these models is a bit like G W Bush looking into Putin's eyes and getting a "sense of his soul".
- lupire 4y ago> G W Bush looking into Putin's eyes and getting a "sense of his soul" That wasn't a mistake. That was Bush being a lying politician. A better example is when police guess whether a suspect is guilty.