I'm thrilled to announce the THINGS initiative: An initiative of researchers around the world collecting and sharing large-scale behavioral and neuroscience data for object recognition and understanding, using the same image dataset.
Ok, since
#IchbinHanna
is trending, for everyone who doesn't know the German academic system and how messed up it is because of its weird laws and rules, here's a short version. 1/n
I just accepted an offer as an independent research group leader at Max-Planck-Institute for Human Cognitive & Brain Sciences
@MPI_CBS
in Leipzig! Very excited to start my own lab! Iโll soon be looking for PhD students and postdocs interested in vision, computation & cog neuro!
Very happy to announce THINGS-data, three massive datasets of fMRI, MEG and similarity ratings of thousands of object images. THINGS-data offers a uniquely broad, systematic and multimodal sampling of object space. For early access please reach out!
I'm very happy to announce that our paper "Revealing the multidimensional mental representations of natural objects underlying human similarity judgements" has now been published in
@NatureHumBehav
.
What did we learn from this work? A thread 1/n
I'm thrilled that THINGS-data is now online! We provide 3 massive datasets of fMRI, MEG, & behavior in response to up to 1854 objects and >22k images. We hope this will allow studying objects in vision, memory & language with unique semantic breadth! Link:
I guess itโs now official: I've accepted a faculty position at University of Gieรen
@jlugiessen
! Iโm eternally grateful to everyone who has supported me on this journey! I will continue my lab
@MPI_CBS
in Leipzig for the next 2 years and will soon start building a team in Gieรen!
๐I'm thrilled and honored to be awarded an
#ERCStG
to work on unraveling the core representational dimensions of ventral visual cortex
@MPI_CBS
. This wouldn't have been possible w/o such a supportive research community and my extraordinary team! Next 5+ years will be exciting!๐คฉ
What are the dimensions underlying our mental representations of objects? We collected ~1.5 million behavioral responses in a triplet odd-one-out task for 1,854 natural objects and developed a computational that identified 49 interpretable dimensions.
Over the last few years, I've been running a number of experiments on Amazon Mechanical Turk. I wrote a guide on some of the things I learned on the way that I would like to share. Hope this is useful.
Feel free to add comments!
But: if you look at permanent admin jobs in academia, they strongly outweigh permanent academic jobs. We need a much broader set of permanent positions, including positions for those who don't want to run labs but who just like to keep doing what they have been doing. 14/n
THINGS: A database of 1,854 object concepts and more than 26,000 naturalistic object images
All concepts, images, metadata (e.g. 27 empirically-determined high-level categories) available here:
THREAD 1/7
Very happy to announce that the preprint for our object concept and object image database is now online! We systematically sampled a set of 1854 object concepts & identified 26107 high-quality images of them (>=12 examples per object), all freely available for academic purposes!
I am looking for two postdocs to join us at Giessen University
@jlugiessen
, in collab. with the Max Planck Institute
@MPI_CBS
in Leipzig, funded by
@ERC_Research
&
@ProLOEWE
. We have exciting projects lined up but also offer a lot of freedom with implementing own ideas!๐ pls RT!
๐ฅณ Iโm happy that our paper โFeature-reweighted representational similarity analysis: A method for improving the fit between computational models, brains, and behaviorโ is now published in
@NeuroImage_EiC
. Why would you want to use FR-RSA? 1/n
Thrilled that our work on object dimensions made it to the cover of
@NatureHumBehav
! The image is a 3D-MDS plot of object dissimilarity for 1854 objects, with the lines reflecting the largest pairwise similarity and the colored dots the most prominent dimension of the object.
I'm seeing more & more people use partial correlation during model comparison to identify the contribution of a model to prediction *after controlling* for other models. However, the correct approach is *not* to use partial correlation. A brief explanation. ๐งต 1/n
In
@MATLAB
, we often iteratively write into new variables. When you don't know the variable's size, you can still preallocate memory by simply inverting the loop, e.g.
clear tmp
for i = n:-1:1
tmp(:,i) = ...;
end
This will then allocate memory in the first call.
#matlabtips
What is the solution? There is no solution that will make everyone happy because academia is a pyramid scheme. There ain't going to be permanent contracts for everyone. 13/n
P.S.: We need to up tenure-track for older academics by adjusting state laws. If you want to apply for a tenure-track job even though you are 40, that should be your choice, and it should be the universityโs choice to say โyesโ or โnoโ.
Can we reconstruct images from brain activity? The
@ykamit
lab argues that recent work uses a non-diverse dataset, and with appropriate controls, performance breaks down completely. This has important implications for what we can learn from such datasets.
At the very least, we need to offer academics the option to decide for themselves if they want to leave academia or stay, which is why getting rid of WissZeitVG would be a great first step! Hence
#IchbinHanna
15/15
I often see people struggle with tiny, unreliable effects in their multivariate neuroimaging studies, be it multivariate decoding or representational similarity analysis. Is there any effect at all? There's a simple strategy to boost your signal: average, average, average. ๐งต 1/7
We need to reduce the incentives for mega-labs that lead to diminishing returns. We need to increase funding but become more selective with training people towards academia, simply by paying PhD candidates more. 15/n
Iโm hiring! If you are interested in doing a PhD or postdoc in computational cognitive neuroscience of visual object recognition, please apply! We offer an excellent and supportive research environment in a fun and dynamic group.For details, see Please RT!
What it also means: If you want to succeed in academia in Germany, choose your postdoc position extremely wisely and perform really well. You literally only have one shot. Or move to another country. 12/n
The German Ministry of Education replies to
#IchbinHanna
by lecturing academics that there actually is a purpose behind the 12 year employment ban. As if they are too naรฏve to know there canโt be permanent jobs for everyone, and ignoring the ๐ฉ๐ชbrain drain.
This is when "Junior" professorships (I kid you not!) were introduced, which were, ironically, fixed-term contracts with no option for tenure. Effectively, it meant hiring postdocs for teaching (but, hey, at least they can apply for 3rd party grants, that's another story). 5/n
I'm soon going to be looking for a senior postdoc to join my lab who is interested in transitioning to independence and in building their own group, with a strong background in vision science, cognitive computational neuroscience or visual neuroscience. Please RT 1/3
When research is correlational, no causal link is found with Mendelian analysis, the effect size is small, the significance statement interprets absence of evidence as evidence of absence, and the author makes health-relevant recommendations on Twitter... ยฏ\_(ใ)_/ยฏ
Do you need a lot of sleep? We find 6.8 hours associated with "optimal" brain structure. Even less for hippo. The recommended 7-9 clearly too much. >8 worse than <5. 51.000 MRIs used. Short sleep not a societal brain health problem?
@LifebrainEU
@LCBC_UiO
We'll soon be looking for a postdoc to join our team! We have exciting projects lined up with densely-sampled fMRI at 7T! A passion for visual neuroscience, fun working in a team, strong fMRI experience, & good coding skills are useful. Starting date flexible. โ๏ธ/DM me & pls RT!
Germany has strict employment laws. With the growing problem of having fixed-term contracts in academia that are extended over and over again and the insecurity associated with this, there is a law in place that puts a cap on how long you can be on non-permanent contracts. 2/n
We are seeking a postdoc at the Max Planck Institute for Human Cognitive and Brain Sciences
@MPI_CBS
for an fMRI project central to our newly funded ERC starting grant COREDIM. Please RT! For details, read on! ๐งต
Happy to announce a new preprint by fantastic PhD candidate Philipp Kaniuth (not on Twitter)! We demonstrate the power of feature-reweighted RSA (FR-RSA) for improving representational similarity analysis. ๐งต
We tried making extracting neural network activations easier for you!
Here is a Google Colab that you can run to get activations for your favorite network, based on our recently published THINGSvision library.
Pytorch:
Tensorflow:
@PessoaBrain
I once recorded a 2 day workshop I gave at NIH on MVPA, with lecture slides:
Other than that, a shameless plug:
Finally, I really like
@haxbylab
โs paper:
Our paper accompanying the Python library THINGSvision for streamlining the extraction of neural network activations has been published at Frontiers in Neuroinformatics! ๐ฅณ
The library can be found here:
What can you do with it?1/5
Are you interested in understanding what dimensions shape our mental representations of objects? Hear me talk
@neuromatch
2.0 on our computational model of object similarity that yields interpretable object dimensions. 2pm EST (8pm GMT+2):
#neuromatch2020
Since almost nobody wants to give you a permanent position if you aren't a professor, this means you have 12 years to be tenured, or you're out. (the time is longer if you are employed on projects, and, for US people: until recently, there was no tenure track) 3/n
Now, a large number of positions are indeed announced as tenure-track, but many of these positions existed already - from retirees! And - surprise - the initial positions are still junior professorships, which means: you cannot apply if you have been a postdoc for >4 years. 10/n
Are you interested in learning about multivariate pattern analysis, firmly embedded in Matlab, and don't know where to start? Check out these great tutorials and YouTube videos by
@AndysBrainBlog
using The Decoding Toolbox (TDT) as an example:
Now, the catch with junior professorships, beyond not offering tenure / increasing teaching load, is: you must not have been a postdoc for more than 4 years (in parts of Germany <6 years since the start of your PhD). Why? Because this would conflict with the 12 year law! 6/n
โก๏ธ THINGSvision just became lightning fast! โก๏ธ
If you would like up to 5x faster extraction times for hundreds of DNN models including ResNet, VGG, and CLIP, check out the updated code, and try it yourself on Google Colab!
Thanks to
@lukas_mut
& co!
Now, a little over 15 years ago, Germany noticed they should offer younger academics an option earlier because such options for permanent jobs are much more attractive in other countries. 4/n
Check out this special issue on โCurrent Topics in Computational Cognitive Neuroscienceโ in Neuropsychologia, edited by
@nico_schuck
and myself. Here is a brief summary of the topic of the special issue and all articles. 1/n
Germany noticed that academic talent was still leaving the country, so a few years ago, the tenure-track initiative was started: There would be rules according to which you would receive tenure after a fixed number of years. Problem solved? 7/n
Our visual world contains ๐ฅ๐๐๐๐๐ค, like plants, people, animals, or vehicles. But there is also ๐ค๐ฅ๐ฆ๐๐, what these things are made of, like glass, metal, or hair.
But how do we make sense of stuff? How do we structure it in our minds? ๐งต 1/n
Of course, this was not so great news for everyone who was working towards a tenured position, because, effectively, an already extremely narrow market just became a lot narrower. 11/n
Ok, so at least 20% of the people I have been following are now active on Mastodon. This is critical mass. Yes it felt clunky in the beginning but Iโm really starting to like it! Distraction with Twitter suggestions/negativity are gone, itโs just science/community. Give it a try!
Ok, ERC Starting Grant submitted. One thing I learned is that no matter how much effort went into writing this and no matter how good it is, there will always be parts that will put off the panel/reviewers, and you just need to hope they'll acknowledge your version of the story.
Ok, this was new to me: in
@MATLAB
you can draw regions in an existing plot (e.g. scatter) and return which of your data points fall inside.
scatter(x,y) % some coordinates
h = drawfreehand;
xi = h.Position(:,1);
yi = h.Position(:,2);
ind = find(inpolygon(x,y,xi,yi);
#matlabtips
Twitter hive mind: how do you best conduct code review in your team in an academic setting? I have been thinking about this more recently but still donโt have a good solution for it. Any help / suggestions appreciated!
Our team just updated our literature overview comparing brains, behavior, and deep neural networks!
Please add missing papers to make it more complete!
(Updates and new list format mostly by
@AnnaWol45981764
, with support of
@hannes_hansen_
)
I just updated the list of papers comparing DNNs, brain & behavior, now including reviews and commentaries. Please add your own papers or other important work if they are missing!
The computationally most efficient way for multiple linear regression in
#Matlab
is:
b = X\Y; % b: regression coefficients, X: design matrix incl. column of 1s, Y: all dependent variables
r = Y-X*b; % residuals
TE = Y-mean(Y);
R2 = 1 - (sum(r.^2) ./ sum(TE.^2)); % expl. variance
Those interested in doing a PhD in psychology or neuroscience: Come and work with us at
@MPI_CBS
in Leipzig! The Max-Planck-Institute offers excellent resources, and Leipzig is as cool, artsy, and affordable as Berlin was 10 years ago.
It turns out that the government had to create incentives for these positions. Why? Because: universities would have to create new permanent jobs, or higher inexperienced people (god forbid!) which may "clog" the professorship for up to 6 years. 8/n
To export editable
#Matlab
figures to e.g. Illustrator, you can often enforce vector graphics by using
hf = gcf;
hf.Renderer = 'painters';
print('im.pdf','-dpdf','-bestfit')
Also, export_fig from File Exchange is very useful!
Check out our new paper in
@eLife
on task and object representations and their interaction. Comes with code and data for running model-based MEG-fMRI fusion.
We often distinguish between objects and places/scenes. But there are things in the world that are both objects and places (e.g. buildings, ships, bridges). Perhaps I'm missing some obvious piece of literature, but has someone looked into how these "place objects" are processed?
Great to see this out! This is the first neuroimaging dataset of the THINGS initiative to be published! ๐
Stay tuned for more news to come soon, check out the other datasets๐, and reach out if you would want to be part of it!
Now published! We recorded EEG ๐ง (n=50) for 22,248 images - all 1,854 concepts in THINGS. We hope this open dataset is useful for understanding how our brains processes and represents objects.
With Ivy Zhou,
@amandarob10
,
@martin_hebart
,
@CompCogNeuro
Interested in extracting features from deep neural networks but not sure where to start? Check out THINGSvision, a library we developed to simplify the extraction of activations from diverse DNNs like AlexNet, VGG, ResNet, CLIP, and Brainscore's
#1
CorNet.
๐จ Preprint alert ๐จThe ViCCo group has its first library! THINGSvision: a Python toolbox for streamlining the extraction of activations from deep neural networks
(1/n)
Here is an introductory lecture on multivariate pattern analysis I recently gave at
@NIMHgov
as part of the fMRI summer course organized by
@fMRI_today
:
Curious to hear what you think!
What do we call the effect that when a project is new, we may find it super exciting but the longer we work on it, the less interesting/novel we find it, even though objectively this hasn't changed?
Want improved results and model selection with RSA? Come and visit poster
#1076
by PhD candidate Philipp Kaniuth
@VSSMtg
! He systematically evaluates the power of feature-reweighted RSA (aka tuned RSA) on a wide range of data sets.
#vss2020
Zoom sessions๐
Hi
#SfN2018
people, come by our posters FF6 and FF7 Sunday afternoon! We identified 49 interpretable behavioral dimensions from similarity judgments for 1,854 object categories. Also, we developed a high-quality large-scale object image database with more than 26,000 images!
This is one of the best ECR opportunities out there. 6+3 years, flexible 2 million EUR, free use of resources (for us: free MRI!), amazing people around you, teaching voluntary (no obligation), very short and easy application procedure. Come and start your own group with us!
Want to head your own independent Research Group?๐ In an exciting research environment & enjoying full scientific & budgetary autonomy? Yes -? Then apply now! ๐คPositions are open for Max Planck Research Group Leaders. Application deadline: May 11, 2023
Nice! Be at a conference, have home made lunch, see your baby, and (well, maybe) work on ethics approval & collect Mturk data - what else could I wish for?
#neuromatch2020
Do you want to build your own research group at our Max Planck Institute in Leipzig, with amazing resources and colleagues? Do you work in computational cognitive neuroscience? Then apply for the call for a Research Group Leader position w/
@doellerlab
๐
With fMRI, it's fascinating to compare results at a group level to individuals. More subtle effects found in most or all individuals can disappear. I'm wondering how prevalent this is, simply because people don't have enough statistical power at the individual level.
@ISDedal
If you get along well with your PhD advisor, they have the funding, and you can continue to be productive, then delay your PhD defense for as long as possible. In most German states, the clock starts ticking only after your PhD.
In my opinion by far the best image reconstruction from brain activity patterns out there. Not only good semantic reconstruction as in many of the current approaches but also nice fidelity of object position and scene layout.
Let's do TLDR for our new preprint! Brain-Diffuser
We reconstructed perceived scene images from brain signals. These reconstructions capture both layout and semantic information of images... But how?
Jack Gallant
@gallantlab
: three different views based on what understanding is based on physics, cognitive science (information processing) and decoding/prediction. Deep nets maybe not consistent with first two views of understanding.
#ccn2018
#ccn18
Outstanding possibility for female early-career scientists to start a research group at a Max Planck Institute, with a really simple and short application procedure. Happy to answer questions about the procedure!
Applications are open for our tenure-track Lise Meitner Research Group Leader positions 2022! The MPG's Lise Meitner Excellence Program is expressly aimed at promoting outstanding female early-career researchers in all areas of
#science
. Apply now!๐
Happy to see this amazing work by Pinglei Bao
@bobking
and
@doristsao
out in
@nature
. They discovered a general organizing principle underlying macaque IT in response to objects, revealing two new object-selective networks & allowing object reconstruction.
Doing something that has not been done before can often yield valuable insights. But the fact that something has not been done before in and of itself is in many cases not a good motivation for a study.
If I have large neuroimaging and behavioral datasets that I would like to share (say fMRI, MEG, ECoG, monkey, behavior), what is the best place to put them? I really like
@OpenNeuroOrg
but is it only so slow for me? What about
@figshare
? Other well known places that I'm missing?
Today I was contacted by a music researcher regarding my high school thesis on overtone singing that I wrote 20 years ago. I happened to be in contact with composer Karlheinz Stockhausen and our conversation was the only literal proof that he invented it independently. 1/2
There seem to be a lot of misunderstandings regarding Amazon Mechanical Turk
@amazonmturk
that keep appearing here on Twitter. Here is my attempt at clearing some of them up. A thread. 1/n
Cultural difference: in the US, it appears to be perceived as ethical and is often required to collect information about race, while in Germany most would be outraged about this question, for obvious historical reasons.
New work by Iris Groen from the
@Chris_I_Baker
lab:
"Distinct contributions of functional and deep neural network features to representational similarity of scenes in human brain and behavior"
w/
@MGreenePhD
,
@ChrisBaldassano
,
@drfeifei
, Diane Beck
People at
@VSSMtg
, please come to my talk Monday morning at 8:15am in room 2! I will be talking about the core dimensions underlying mental representations of objects and how we identify them using large-scale behavior and computational modeling.
#VSS2019