Blake Richards / is on Nostr

Researcher in Montréal combining artificial intelligence and neuroscience. I post mostly science, occasional politics and house music.

Last Notes

2023-11-17 19:49:37

- reply

@npub19d9…f02p @npub1pe4…rjva @npub15sw…x855
I should note one thing after a bit more reflection: as someone who did their grad and postdoc work in experimental labs, and who has engaged with a lot of experimentalists, I actually don't think they're always better on this front necessarily...
I think there are more than enough experimentalists who ignore any confounds that don't actively stop them from collecting data. 😅

2023-11-17 16:12:34

- reply

@npub19d9…f02p @npub1pe4…rjva @npub15sw…x855
I will say that I can admit that we don't always do as many controls as we should - it's sometimes just too much compute!

2023-11-16 22:12:22

- reply

@npub15sw…x855 @npub19d9…f02p
But this is what I'm saying: people are doing the work to explore what leads to good fits and what doesn't, and it's not simply "because there's lots of parameters". The real answer seems to be more complicated.
Also, I have to note that this result you've posted here is the opposite of that found by a different recent paper, where they found that better next word prediction did lead to better matches:
https://www.biorxiv.org/content/10.1101/2020.07.03.186288v2.abstract
So, TBD on this particular Q...

2023-11-16 21:47:07

- reply

@npub15sw…x855 @npub19d9…f02p
2. The hard work you describe is precisely what people are trying to do, no? Like, concretely, what is it you think is not being done that should be?

2023-11-16 21:46:22

- reply

@npub15sw…x855 @npub19d9…f02p
I'm not sure I follow either of you...
1. The number of parameters is immaterial, because you're not fitting *the data*, you're fitting a different task, then asking whether the model arrives at a solution that fits the data. So the number of parameters is not helping you with that fit for trivial reasons.

2023-11-16 21:02:42

- reply

@npub15sw…x855
I don't agree. The point is that you construct different models based on different hypotheses and then evaluate which ones produce better fits. So, you very much so learn something.

2023-11-16 20:10:27

- reply

@npub15sw…x855
The data is a huge driving force, but many papers compare different losses and architectures, and show that some match the brain better than others.

2023-09-11 16:31:30

- reply

@npub1skv…aky3 @npub1qe8…4vky @npub17p4…zzd8 @npub1vvs…k2yk @npub1pe4…rjva @npub19un…gacg
Fascinating discussion! Thanks for sharing this.
But, I can't help myself, I have to engage in my usual refrain here (sorry, lol), since it comes up in the second paragraph:
Brains are *literally* computers and information processing devices, it's not a metaphor!!! 🙂

2023-08-30 18:48:06

- reply

@npub1skv…aky3 finally got a chance to finish reading this. I agree with it so much.
In particular, the "pick a noun" paradigm is something that has bugged me for ages. It will be hard for psychologists to kick that habit though! It's a paper generating machine...