Sign in

Alexander Huth

@alexanderhuth.bsky.social
1.7K followers 306 following 92 posts

Interested in how & what the brain computes. Professor in Neuroscience & Statistics UC Berkeley

PostsRepliesMedia
Reposted by Alexander Huth
jerrytang.bsky.social @jerrytang.bsky.social · 08/09/2026
Our new study on cross-participant cortical mapping (with @alexanderhuth.bsky.social) is out in @imagingneurosci.bsky.social! direct.mit.edu/imag/article... 1/7
1175
Reposted by Alexander Huth
Cogan Lab @coganlab.bsky.social · 21/07/2026
Last week, Areti Majumdar (2nd year PhD student) presented @jerrytang.bsky.social and colleague’s seminal paper on reconstructing continuous language from non-invasive brain recordings. This 🧵 explores our thoughts and questions.
162
Reposted by Alexander Huth
Alison Preston @aliprestonphd.bsky.social · 29/05/2026
Proud mentor moment! My student @mujianing.bsky.social (co-supervised w/ @alexanderhuth.bsky.social) presents at #CEMS2026 today. Her work challenges the idea that event boundaries are remembered because they're surprising. What predicts memory is how much a moment shares with the rest of the story.
0273
Reposted by Alexander Huth
Ole Goltermann @olegolt.bsky.social · 20/05/2026
A recent paper by Epp et al. in NN claimed that ~40% of reported BOLD findings could be misinterpreted. In our reanalysis, we identified several statistical issues that, in our view, undermine these conclusions. www.biorxiv.org/content/10.6... 👇 1/11
biorxiv.org
14626
Alexander Huth @alexanderhuth.bsky.social · 09/04/2026
Really excited about our new work on aphasia! Even in fairly profound aphasia, we can recover semantic maps through visual stimuli and use them to decode language. This is a big step! Language BCIs in aphasia might be possible!
06817
Reposted by Alexander Huth
Jack Gallant @gallantlab.org · 23/03/2026
Pycortex, our awesome brain viewer software, has a major new feature! Thanks to the hard work of Dr. Matteo Visconti and Aditya Vaidya, Pycortex now supports headless webgl viewers! You can create 3D inflated plots without needing a display, and in Jupyter notebooks! gallantlab.org/pycortex/aut...
gallantlab.org
Plot 3D Brain Views Headlessly (No Browser) — pycortex 1.3.0.dev0 documentation
0235
Alexander Huth @alexanderhuth.bsky.social · 07/01/2026
I was spurred to respond to your paper by the fact that it's being shared widely outside neuroscience. It raised the specter of the dead salmon paper (the number of times I've heard about that damn salmon as a gotcha from non-neuroscientists..), so I wanted to dig in and understand the results.
0180
Alexander Huth @alexanderhuth.bsky.social · 07/01/2026
Finally, you may not consider the 40% number the "headline result" of the paper but it is (1) in the abstract, and (2) the literal headline for news stories about this paper. Broadcasting the idea that most fMRI results are hogwash may not be your intent, but it is definitely the outcome
180
Alexander Huth @alexanderhuth.bsky.social · 07/01/2026
When we repeatedly show fixating subjects short movie clips, these areas often show negative BOLD correlation between repetitions! But with longer and more engaging stimuli the correlations flip to positive. Real functional nonstationarity is anatomically localized and could contribute strongly here
120
Alexander Huth @alexanderhuth.bsky.social · 07/01/2026
I agree with you that the spatial clustering is a strong signal, and replicating it makes it stronger. But the localization of discordant voxels to VAN/DMN makes me uneasy. These are the areas where we see the most nonstationarity in BOLD as subjects' attention wanders in and out during experiments
220
Alexander Huth @alexanderhuth.bsky.social · 07/01/2026
The flip from +ve to -ve correlation btwn OEF and CMRO2 in discordant v concordant voxels (your Fig. 5d) is actually also replicated by this simulation — I think it emerges naturally from selecting based on discordance.
110
Alexander Huth @alexanderhuth.bsky.social · 07/01/2026
I am less confident in the assumed noise levels, which seem plausible but are not empirically based. Still, the only noise that seems to matter is that in the CBF measurement, and its scale here (stdev=1 ml/100g/min) is not crazy, I think.
110
Alexander Huth @alexanderhuth.bsky.social · 07/01/2026
Thank you for your responses, Valentin! On this: I tried to match the contrast values to your results — the "true" ΔCBF is based on your reported 7.7% from Table S1; others are from Figure 2b,c. The noise-free simulation shows ΔCBF ranging from -7..7%, well within the -15..30% range in your Fig. 3b.
2132
Alexander Huth @alexanderhuth.bsky.social · 06/01/2026
yes, that is a stronger signal. But all of the underlying datasets are also spatially autocorrelated, which will induce some clustering of the results. I also would not be surprised if noise level in the CBF measurement varied across the cortical surface, which would cause consistent clustering
020
Alexander Huth @alexanderhuth.bsky.social · 06/01/2026
I totally agree that negative BOLD is weird and probably not indicative of decreased metabolism. That's the part of this recent paper that I found most compelling.
110
Alexander Huth @alexanderhuth.bsky.social · 06/01/2026
yes absolutely. I simulated uniform "true" activations symmetric about zero, but the real data has more +BOLD than -BOLD.
000
Alexander Huth @alexanderhuth.bsky.social · 06/01/2026
yes, I think it depends on the distribution of "true" values. I used uniform values symmetric about zero in the simulation, so the % discordant is equal for + and -. Shifting the distribution to have more positive values than negative (realistic) makes more of the negatives discordant.
010
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
In short, I think the headline result is wrong, or at least statistically unsupported. The results are consistent with BOLD perfectly tracking CMRO2. (I'm not saying it does, but I don't think the data says it doesn't.)
4150
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
My simulation is here if anyone wants to play with it: github.com/alexhuth/not...
github.com
notebook-sharing/bold-cmro2-cbf-r2.ipynb at master · alexhuth/notebook-sharing
Contribute to alexhuth/notebook-sharing development by creating an account on GitHub.
1182
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
The statistically nasty thing here is that CMRO2 is a vey derived metric, so noise in one part affects all. It turns out you only need noisy CBF measurement to get the discordance effect, because CMRO2 is a product of CBF and some other factors.
160
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
So: simulating the data-generating process with no discordance but realistic noise gives exactly the same result as the paper (40% discordance). I think this means the result does not exclude the null hypothesis that BOLD and CMRO2 always covary positively.
1282
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
But (shocker) with realistic amounts of noise you recover something that looks exactly like their plot. It even has roughly the same level of discordance, about 40% in total!
4475
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
So I made a little simulation where there is no actual discordance, i.e. the underlying values from which the CBF, CBV, BOLD, and T2* are measured all move in lockstep. With zero noise this gives you zero discordant voxels (nice little diagonal line).
1111
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
(Also the measures here are not that simple: ΔCMRO2 in particular is an extremely derived metric that includes both BOLD and CBF!)
170
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
My gut check: that’s a big noisy blob, and they’re saying all points have to fall in those two triangles? Zero chance with real data. Each measure is noisy, and they’re making hay out the fact that if you stratify on one of the measures, the other is not also perfectly stratified
1130
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
They go on to claim that ~40% of all voxels are discordant, including ~25% of voxels with +BOLD and ~60% of voxels with -BOLD. The 40% figure is in the abstract.
150
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
The key result is Figure 3b, which shows ΔCMRO2 vs ΔCBF vs ΔBOLD for a bunch of voxels. Voxels that fall (1) to the right of ΔCMRO2=0 and below the diagonal or (2) to the left of 0 and above the diagonal are called “discordant” because their ΔBOLD and ΔCMRO2 have different signs
161
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
They also include several other measures, including especially CBF (cerebral blood flow), which is used to compute CMRO2.
130
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
CMRO2 is compared to the more classic BOLD (blood-oxygen-level-dependent) signal, which is easy to measure but not quantitative and complex in origin (it involves changes in blood oxygen, blood volume, and blood flow).
180
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
The paper is based on the quantitative functional metric CMRO2 (cerebral metabolic rate of oxygen), which is a calibrated measure of how much oxygen is consumed in a piece of brain tissue during some period of time.
190
Alexander Huth @alexanderhuth.bsky.social · 05/01/2026
This paper had a pretty shocking headline result (40% of voxels!), so I dug into it, and I think it is wrong. Essentially: they compare two noisy measures and find that about 40% of voxels have different sign between the two. I think this is just noise!
8241100
Reposted by Alexander Huth
Neuron @cp-neuron.bsky.social · 17/12/2025
dlvr.it
Linguistic coupling between neural systems for speech production and comprehension during real-time dyadic conversations
Zada et al. use fMRI hyperscanning to simultaneously record dyads engaging in free-form, interactive conversations. They find that speech production and comprehension rely on highly overlapping neural representations across the cortical language network. Brain-to-brain coupling is strongest in areas associated with social cognition.
0124
Alexander Huth @alexanderhuth.bsky.social · 12/12/2025
the world is ~1/f. therefore the brain that perceives and models the world should be ~1/f. 🤷‍♂️
250
Reposted by Alexander Huth
Rodrigo Braga @rodbraga.bsky.social · 07/10/2025
📣 New preprint from the Braga Lab! 📣 The ventral visual stream for reading converges on the transmodal language network Congrats to Dr. Joe Salvo for this epic set of results Big Q: What brain systems support the translation of writing to concepts and meaning? Thread 🧵 ⬇️
26117
Alexander Huth @alexanderhuth.bsky.social · 01/10/2025
Always include some stimuli that are permissively licensed so they can be used as examples! E.g. we have a video stimulus set that's mostly Pixar short films, but also includes a segment from the Blender movie Sintel (en.wikipedia.org/wiki/Sintel), which is licensed CC-BY.
en.wikipedia.org
Sintel - Wikipedia
030
Alexander Huth @alexanderhuth.bsky.social · 13/09/2025
Happy and proud to see @rjantonello.bsky.social’s work awarded by SNL!
1304
Reposted by Alexander Huth
Alejandro de la Vega @neurozorro.bsky.social · 11/09/2025
Our latest paper outlining our ecosystem of tools for mining the neuroimaging literature, is finally officially published in eLife! doi.org/10.7554/eLif...
doi.org
Mining the neuroimaging literature
New tools for literature mining, such as automated analysis of the research literature, are accessible, scalable, and reliable.
1204
Alexander Huth @alexanderhuth.bsky.social · 10/09/2025
Someone on the aggies subreddit posted the course description and it seems pretty thorough! www.reddit.com/r/aggies/s/C...
reddit.com
Wang_Lung_1921's comment on "Regardless of your stance on the politics behind the Welsh/LGBT situation…"
Explore this conversation and more from the aggies community
050
Reposted by Alexander Huth
Heejung Jung @jungheejung.bsky.social · 04/09/2025
New Open dataset alert: 🧠 Introducing "Spacetop" – a massive multimodal fMRI dataset that bridges naturalistic and experimental neuroscience! N = 101 x 6 hours each = 606 functional iso-hours combining movies, pain, faces, theory-of-mind and other cognitive tasks! 🧵below
411658
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
And of course check out the paper! www.biorxiv.org/content/10.1...
biorxiv.org
Evaluating scientific theories as predictive models in language neuroscience
Modern data-driven encoding models are highly effective at predicting brain responses to language stimuli. However, these models struggle to explain the underlying phenomena, i.e. what features of the...
020
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
or @csinva.bsky.social: bsky.app/profile/csin...
100
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
I'm posting this thread to highlight some things I thought cool, but if you're interested you should also check out what @rjantonello.bsky.social wrote: bsky.app/profile/rjan...
110
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
Cortical weight maps were also reasonably correlated between ECoG and fMRI data, at least for the dimensions well-captured in the ECoG coverage.
100
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
Finally, we tested whether the same interpretable embeddings could also be used to model ECoG data from Nima Mesgarani's lab. Despite the fact that our features are less well-localized in time than LLM embeddings, this still works quite well!
100
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
To validate the maps we get from this model we also compared them to expectations derived from NeuroSynth and results from experiments targeting specific semantic categories, and also looked at inter-subject reliability. All quite successful.
100
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
The model and experts were well-aligned, but there were some surprises, like "Does the input include technical or specialized terminology?" (32), which was much more important than expected.
100
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
This method lets us quantitatively assess how much variance different theories explain about brain responses to natural language. So to figure out how well this aligns with what scientists think, we polled experts to see which questions/theories they thought would be important.
100
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
"Does the input include dialogue?" (27) has high weights in a smattering of small regions in temporal cortex. And "Does the input contain a negation?" (35) has high weights in anterior temporal lobe and a few prefrontal areas. I think there's a lot of drilling-down we can do here.
100
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
The fact that each dimension in the embedding thus corresponds to a specific question means that the encoding model weights are interpretable right out-of-the-box. "Does the input describe a visual experience?" has high weight all along the boundary of visual cortex, for example.
Left hemisphere cortical flatmap showing regression weights for the feature "Does the input describe a visual experience or scene?"
100
Alexander Huth @alexanderhuth.bsky.social · 18/08/2025
But the wilder thing is how we get the embeddings: by just asking LLMs questions. Each theory is cast as a yes/no question. We then have GPT-4 answer each question about each 10-gram in our natural language dataset. We did this for ~600 theories/questions.
100