Jayson Jeganathan

Do you work on surface MRI such as HCP data? Our paper is finally out in Imaging Neuroscience, showing biases in surface fMRI. The revised ms has many new insights! TLDR: Vertices are much closer to each other in sulci, resulting in (fake) high spatial autocorrelation. (1/5)

@alexblania @grok what is this

World is now available in the USA.

New paper in Imaging Neuroscience by Jayson Jeganathan, Michael Breakspear, et al: Spurious correlations in surface-based functional brain imaging doi.org/10.1162/imag_a…

@GianDuma This specific issue of mapping onto an irregular surface mesh is likely not relevant to MEEG. But you can have a different gyral bias in MEEG as you outlined

@jaysonjeg Fantastic work, congrats! In your opinion, does this apply to MEEG source activity too? In source activity, specially in EEG, sulci activity is less represented, therefore I am not sure to expect the same behavior of fMRI. Have you any data about it?

Do you work on surface MRI such as HCP data? Our paper is finally out in Imaging Neuroscience, showing biases in surface fMRI. The revised ms has many new insights! TLDR: Vertices are much closer to each other in sulci, resulting in (fake) high spatial autocorrelation. (1/5)

This was a team effort! @KoussisNikitas @tremodian @Sina_Mansour_L @AndrewZalesky @DrBreaky @phogat_richa @jchrispang

I'm proud of this work, and I hope this paper helps you to avoid spurious conclusions in your work. Check out the full paper (tinyurl.com/gyral) for a detailed breakdown and for other downstream analyses that can be impacted (FC fingerprinting, hyperalignment, etc). (5/5)

Amazing paper !

In markets, there are price bubbles. In science, are there popularity bubbles? When promising research gets hyped up, it can lead to a burst of activity followed by a dramatic collapse, shows huge bibliometric study @Nature Magazine l.sslevine.com/Ex7z

Wow. Most recent openAI model o3 smashes the prev leaderboard for ARC-AGI reasoning challenge. It is also now the 175th best coder in the world (CodeForces challenge), and can solve 25% of hard-for-humans FrontierMath Qs (up from 2%). What does this all mean?

Hi friends, find me at the other place on jaysonjeg.bsky.social

jaysonjeg.bsky.social I'm on bluesky !

I've combined some of these together into a new #neuro #visualization "In the eye of a fly": Using the @FlyWireNews visual system wiring diagram, #Brian2 spiking #neuron simulator, and #blender. Information flow through the chiasms, across a topographic map.

@PessoaBrain My priors had brains quite close to artificial neural networks. Guess that's why I'm surprised. Also it's astounding that neurons can learn and rewire on such short time frames

@jaysonjeg One way only?? 😮

𝗛𝗼𝘄 𝘀𝗵𝗼𝘂𝗹𝗱 𝗼𝗻𝗲 𝗱𝗼 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗻𝗲𝘂𝗿𝗼𝘀𝗰𝗶𝗲𝗻𝗰𝗲?? This way! Every neuro program should just make this required reading and teach students how to do outstanding science like this. doi.org/10.1038/s41586…

@PessoaBrain Yea this seems to tell us one way that brains work differently from neural networks

@jaysonjeg deep neural networks are a different species right? Why would they behave the same as a rodent brain??

Wow

@DrYohanJohn In original transformer paper, it was for translation. The Wq could be applied to an English token, and the Wk to a Spanish token. In the cross attention layers of multimodal foundation models, Wq and Wk could be applied to text and visual input respectively

@DrYohanJohn It's been much nicer having this kind of detailed science discussion than the usual Twitter storm and doom scrolling :)

I've been trying to wrap my head around the transformer mechanism, but something about this "key, query, value" business eludes my grasp. What is the value exactly? I understand the math, but I don't see the purpose of it. Thoughts?