Kai Sandbrink

24 posts

Kai Sandbrink

Kai Sandbrink

@akaijsa

Computational cognitive neuroscience PhD Student @UniofOxford & @EPFL_en co-supervised by @summerfieldlab and @compneuro_epfl

Katılım Mart 2023
179 Takip Edilen155 Takipçiler
Kai Sandbrink
Kai Sandbrink@akaijsa·
Excited that my paper on metacontrol in humans and neural networks with @summerfieldlab and @LHuntNeuro is out in PNAS! We examine the way that predictive representations of control enable behavioral adaptation across settings, and links to pathologies: pnas.org/doi/10.1073/pn…
English
3
13
55
4.9K
Kai Sandbrink
Kai Sandbrink@akaijsa·
This was an amazingly fun collaboration that began at the AC Summer School @GatsbyUCL, and I can't wait to see where it takes us next! Check out our spotlight paper and poster at NeurIPS. Grainy picture from presenting the first version of this work at @CosyneMeeting attached!
Kai Sandbrink tweet media
Alexandra Proca@a_proca

Thrilled to share our NeurIPS Spotlight paper with @japhba* @akaijsa* @SaxeLab @summerfieldlab @Ahummos*! We study how task abstractions emerge in gated linear networks and show that these abstractions support cognitive flexibility. openreview.net/pdf?id=AbTpJl7…

English
1
1
21
2K
Kai Sandbrink
Kai Sandbrink@akaijsa·
It was a pleasure teaching at the Computational Psychology and AI Summer Schools for undergraduate students. Thanks to @lmhoxford @UniofOxford for hosting! @VedanthNath, it was great having you in the lectures - I look forward to seeing what you are up to in the future!
Ved@VedanthNath

I'd have never thought that'd I'd get into Computational Neuroscience AND really like all Computing/Mathy part of it. Thank you to the most amazing people at @summerfieldlab for exciting me about this space, especially Dr. @akaijsa!

English
0
0
7
643
Ved
Ved@VedanthNath·
I'd have never thought that'd I'd get into Computational Neuroscience AND really like all Computing/Mathy part of it. Thank you to the most amazing people at @summerfieldlab for exciting me about this space, especially Dr. @akaijsa!
Ved tweet media
English
1
0
4
760
Kai Sandbrink
Kai Sandbrink@akaijsa·
Remarkably, we find that this individual variation in behavior correlates well with PCs extracted from anxiety & depression and compulsivity transdiagnostic factor scores. We hope these findings can pave the way for using ANNs to study healthy and pathological meta-control! (4/4)
Kai Sandbrink tweet media
English
1
0
2
244
Kai Sandbrink
Kai Sandbrink@akaijsa·
We perturb the hidden representations of the meta-RL networks along the axis used for APE prediction. When perturbed systematically, the models replicate human individual differences in performance across levels of controllability (3/4)
Kai Sandbrink tweet media
English
1
0
2
325
Kai Sandbrink
Kai Sandbrink@akaijsa·
Excited that the preprint for the work from my first two years of PhD at @summerfieldlab is out! In this work, we examine the role of action prediction errors (APEs) in cognitive control: osf.io/5ezxs (1/4)
English
2
13
72
6K
xuan (ɕɥɛn / sh-yen)
xuan (ɕɥɛn / sh-yen)@xuanalogue·
@akaijsa @summerfieldlab One could also use these tools to look for the neural signatures of known planning algorithms, in order to test whether people or ANNs have learned to approximate such algorithms: twitter.com/xuanalogue/sta…
xuan (ɕɥɛn / sh-yen)@xuanalogue

Belatedly discovered this paper yesterday! Really cool to find out that the algorithms I've been using to model human planning -- depth-limited & heuristic-guided forward search (A*, RTDP / RTHS, MCTS) have some neuroscientific evidence for them!

English
1
0
1
173
Kai Sandbrink
Kai Sandbrink@akaijsa·
@xuanalogue @summerfieldlab However, I'm curious how you would operationalize the difference: What kind of predictions do you think one would make if (a) holds over (b), or vice versa?
English
1
0
1
131
Kai Sandbrink
Kai Sandbrink@akaijsa·
@xuanalogue @summerfieldlab Thanks, Xuan! That's an interesting link. Personally, I think it's possible that both are true, that you can both learn certain planning-type mechanisms through trial and error as well as that trial-and-error learning can result in behavior that looks as though it is planned!
English
1
0
2
121
Kai Sandbrink
Kai Sandbrink@akaijsa·
The final work is ours: deep RL models of meta-control. By learning predictive representations of their own control, these networks distinguish situations requiring different in-context updates. We presented this work at CCN 2023 and will release a longer paper soon - stay tuned!
Kai Sandbrink tweet media
English
0
0
4
593
Kai Sandbrink
Kai Sandbrink@akaijsa·
Needing no introduction, the second work we review is deep meta-RL. Along with large language models, these architectures have the ability to learn solutions that adapt to new situations in-context, effectively "learning-to-learn" (3/4)
Kai Sandbrink tweet media
English
1
0
2
875
Kai Sandbrink retweetledi
Sophia Becker
Sophia Becker@sobeckerneuro·
Wondering how representational similarities influence our perception of novelty and novelty-driven behaviors? Then don't miss my talk this Friday in #Cosyne2024! Looking forward to great questions and discussions! 🤩😊
Sophia Becker tweet media
English
1
9
37
2.9K
Kai Sandbrink
Kai Sandbrink@akaijsa·
I had a great time at the @CIFAR_News #DLRL summer school last week! What a fantastic opportunity to learn about cutting-edge methods in ML and explore the fantastic research environment at @Mila_Quebec
Kai Sandbrink tweet media
English
0
0
4
289

Keşfet

@summerfieldlab @LHuntNeuro @GatsbyUCL @CosyneMeeting @lmhoxford @UniofOxford @VedanthNath @akjagadish