Professor, Department of Psychology and Center for Brain Science, Harvard University
Cambridge, MA
- Tweets 2,686
- Following 58
- Followers 16,560
- Likes 5,150
p-hacking: when a child asks multiple parents for permission until they get a positive result.
45
1,032
8,622
My 3-year-old daughter said she wanted to be a professor, so I asked her what professors do. "Talk to each other, drink wine, and put trees on their heads." Sounds about right.
14
154
1,280
Everyone racking their brains about explainable AI should have a look at the cognitive science literature on how people explain their own behavior. People have very detailed explanations. The only problem is that these explanations can be very wrong!
41
147
1,064
Coming in October!
What Makes Us Smart: The Computational Logic of Human Cognition
press.princeton.edu/books/pa…
13
115
768
I was thinking about writing a paper called "What the hell is the insula doing?" but then realized I could write this paper about literally any brain region.
14
42
650
I was asked to present some advice to young investigators. Here are my slides. Happy to hear suggestions!
gershmanlab.webfactional.com…
25
171
588
As a reviewer for NeurIPS, I'm surprised that no reviewers want to give high ratings to any papers. Even those I gave an 8 or 9 to, nobody else gave higher than 6. Why is everyone so negative? A good paper doesn't have to be perfect as long as it's making progress.
18
29
559
Before you write a review of a paper in which you say "It's mostly meh" (a direct quote), stop to consider that you might be talking to a grad student who spent years on the project and is now internalizing the social norms of their profession.
25
42
521
57,006
Nature Communications charges $5890/article and has published 159 articles in the last 7 days. Much of that is paid by federal grants. This means that the money spent on *one* journal over the course of 2 weeks could support an entire graduate cohort for a year.
8
130
464
Wired is using something called "hype", a concept from quantum mechanics, to confuse familiar concepts with cutting-edge research.
Stitch Fix is using something called eigenvector decomposition, a concept from quantum mechanics, to tease apart the overlapping “notes” in an individual’s style. Using physics, the team can better understand the complexities of the clients’ style minds. wired.trib.al/IvWQkYM
4
48
495
I told my 3yo that anything can break with enough force and he asked me if holes can break.
19
11
444
Needless to say, I told my kids to preregister their design and correct for multiple parents.
6
22
458
People sometimes ask me about reading suggestions for cognitive science and related topics, so I thought I would put together a (highly subjective) annotated reading list: gershmanlab.webfactional.com…
I might expand this later since I'm sure I forgot some things.
10
124
450
I don't understand "data available upon request". Isn't that actually more work for the authors? Wouldn't it be easier to document it once and put it somewhere on the internet rather than having to respond to e-mails long after you've forgotten where you've even put the data?
32
52
439
I'm looking for examples in the history of neuroscience where no one thought to measure something until a theory posited that it should exist (i.e., empirical phenomena that were effectively invisible in the absence of theory).
58
99
411
What's up with undergrads missing class constantly due to sickness? I've heard this is a thing at multiple institutions. My class has been repeatedly decimated.
782
69
375
1,106,967
I've been spending time analyzing open-source systems neuroscience datasets, and my general impression is that there are major problems with how people organize their data and write their code. This needs to be a bigger part of training.
9
47
335
71,013
Sometimes when a reviewer says your idea is trivial or obvious, it feels a bit like the museum visitor saying "my 5-year-old could have painted that". Well, why didn't they? (Or to paraphrase Grouch Marx, "Fetch me a 5-year-old").
14
37
319
If you're wondering whether neuroscience is useful for AI, consider the counterfactual world where engineers knew nothing about neuroscience. Do you think that we wouldn't have convolutional networks or reinforcement learning?
32
35
335
How significant is the amount of federal funding spent on article processing charges? Let me try to put this in perspective...
5
108
319
Norbert Wiener, who never took breaks, was overworking one of his students, until the student figured out he could turn Wiener off by telling him he wasn't the greatest mathematician in the world, then turn him on again by telling him he was indeed the greatest mathematician.
3
49
301
Possibly unpopular opinion: most science grad programs aren't set up to train students for non-academic paths. We should have more training programs for students who are interested in science but ultimately don't want to become scientists (in the sense of running their own labs).
20
19
277
65,420
Correct. Your brain is actually a pile of unfolded laundry with a tiny seahorse inside eating an almond.
journals.sagepub.com/doi/ful…
3
26
252
I think it's great that the White House is mandating OA, but only if they're not simultaneously mandating massive infusions of research funding into the coffers of corporate publishers. We could use that money instead to support actual research.
11
33
247
What does the free energy principle tell us about the brain?
arxiv.org/abs/1901.07945
I'm eager to get feedback on this, particularly if I've misunderstood or misconstrued anything about the theory.
13
91
250
Very nice article about learning in single cells: the-scientist.com/features/c…
4
70
271
I'm glad to see that Frontiers is taking a detail-oriented, personalized approach to recruiting topic editors. I consider my "Author Correction" paper to be one of my greatest contributions to science.
4
8
243
The @NeurIPSConf reviewing process is absurd. We're asked to read and respond to reviews, replies to reviews, replies to replies, and so on, yet there is no mechanism for submitting a revised paper. That's what matters at the end of the day.
13
10
263
New survey on predictive representations in reinforcement learning, covering AI, cognitive, and neuroscience perspectives:
arxiv.org/abs/2402.06590
Joint work with @cogscikid, @momchiltomov, @willdecothi, @caswellcaswell
69
259
24,775
Something I've been cooking for a while: a synthesis of synaptic and molecular memory mechanisms.
gershmanlab.com/pubs/memory_…
I welcome critical feedback!
17
46
251
Memory as a computational resource:
authors.elsevier.com/c/1cPSa…
Ishita Dasgupta and I weave together several strands of thought from computer science and psychology. As Hofstadter quipped, "Cognition is recognition."
5
52
244
Article in Nature about mass editorial resignations at journals:
nature.com/articles/d41586-0…
If you are a disgruntled cognitive science editor, reach out to me and Ted Gibson (@LanguageMIT). We run Open Mind (direct.mit.edu/opmi). Free to publish, free to read.
8
69
233
51,312
I'm pleased to announce that Open Mind has replaced Editorial Manager with a lovely system designed by @janeway_systems at very low cost. This is part of our efforts to show that a high quality OA journal can be run cheaply, without the need for article processing charges.
5
26
210
As opposed to what??
Popular books and articles about the brain often promulgate a form of Cartesian dualism, making it sound surprising that mental events are produced by brain events.
8
17
220
I've written a little commentary about biologically inspired AI:
gershmanlab.com/pubs/NeuroAI…
Feedback is welcome.
18
40
234
104,829
Just looking: the innocent eye in neuroscience
gershmanlab.com/pubs/innocen…
Looking forward to any feedback!
16
36
217
I hope Harvard does the right thing and uses its new unrestricted $300M to improve the lot of its postdocs and lecturers.
For example, it could subsidize daycare for postdocs with children. It could make all lecturer positions well-paid and indefinitely renewable.
4
14
215
37,136
Our work, led by Pedro Tsividis, on theory-based reinforcement learning: arxiv.org/abs/2107.12544v1
@MITCoCoSci
4
53
219
I joked with someone that if computer scientists wanted to emulate biological intelligence, they should restrict their energy budgets to ~20 watts (the power used by the human brain). But then I pondered this further, and it's not quite correct.
4
27
223
I plan to hire a postdoc next summer to work on machine learning projects at the interface with cognitive science and neuroscience. Part of a collaboration with @pulkitology, @pabbeel, @AudeOliva, and @FieteGroup. Get in touch with me if you're interested.
3
91
222
46,896
We need a cognitive science of scholarship, starting with the recognition that attention is limited: you can't read everything. Historically, we have had journals that function as an attentional filter.
12
19
213
There was a point when I was considering a switch to a research position in industry, and one thing that gave me pause was the remarkable longevity of universities compared to corporations.
13
16
212
Big news: I actually used the quadratic formula to solve a research problem this morning! I never thought this day would come. I'd like to thank all of my middle school math teachers for their unwavering support.
5
1
206
12,036
Habituation is arguably the simplest form of learning, yet surprisingly it remains mysterious. I try to distill some elementary principles here:
osf.io/preprints/psyarxiv/ys…
Feedback is welcome!
3
34
202
33,963
New lab tradition: every time someone gets a paper accepted, we add a component to a Rube Goldberg machine in my office. To animate the joys of uselessness.
2
14
180
4yo (crying): "I'm not having a big feeling!"
7yo: "Yes you are, crying is a feeling."
4yo: "No it's not!"
7yo: "Yes it is. You want dad to look inside your head and check if it's a feeling?"
2
2
187
The professor central limit theorem: if you give a professor enough random tasks to do, their output becomes white noise.
15
201
I'm looking to take a new PhD student interested in studying learning in single-celled organisms, both experimentally and theoretically. Email me if you're curious!
6
71
193
The examples in this article are so wild. If even half of these are true, something truly profound is happening here which we don't understand.
This is a morbid example but, memory transference in transplant recipients is a repeated phenomenon and should be studied more. All of y'all who think replicating human intelligence/memory = recreating a brain in digispace are kinda losing the plot IMHO. pubmed.ncbi.nlm.nih.gov/3173…
15
31
200
50,735
The reward prediction error hypothesis of dopamine has been both influential and controversial. My colleagues (J. Assad, @scott_linderman, @naoshigeuchida, @blsabatini, @WilbrechtL, @Datta_Lab) and I discuss some of the puzzles and what they mean: nature.com/articles/s41593-0…
3
54
194
19,741
I'd like to teach a paper which shows how a fact about the brain materially improved an AI system in a way that is unlikely to have been figured out by engineering alone. I haven't been able to find a single example of this. Suggestions welcome.
41
21
206
63,639
I hope that 300 years from now, in the next pandemic, someone will say "When Harvard was shut down, Sam Gershman repaired to his house and invented a new way to get his kids to eat dinner."
2
2
176
New work on how the brain might solve reinforcement learning tasks with hidden states:
biorxiv.org/content/10.1101/…
with @jehosafet, @naoshigeuchida, @Sandra_RoPi, @scott_linderman
46
180
27,255
Here’s a challenge: is there a strong inference about cognition from neural data that could not more easily, unambiguously and cheaply be obtained from behavioral data?
42
60
171
Just to put this in perspective, the cost of running Open Mind for a year is roughly the total APCs collected by Nature Communications in a single day. Journals can be run inexpensively, without gouging authors (and by extension taxpayers).
The fundamental question facing open access publishing is how it will be funded. Harvard and MIT libraries are taking the lead (in partnership with MIT Press) to directly fund open access journals. This will hopefully expand to more libraries so more journals can be supported.
3
44
177
19,849
Trust me that as a computational neuroscientist, I would very much like to be situated at the center of an AI revolution, but I just don't think it works that way. New engineering ideas come from thinking about the structure of problems, not reading the tea leaves of biology.
10
10
184
fMRI people: please stop referring to the "dopaminergic midbrain" when you just mean "midbrain". There are multiple cell types in the midbrain (not just dopamine neurons), and fMRI does not tell you which are contributing to the BOLD signal.
1
21
166
19,694
When are universities going to establish "History of Artificial Intelligence" departments for teaching papers written before 2012? Those ancient scribes, who left primitive ideas entombed in amber and volcanic ash, may have much to teach us about our past.
4
12
174
Can unicellular organisms learn? We argue yes:
gershmanlab.com/pubs/cell_le…
joint work with Petra Balbi, Randy Gallistel, and Jeremy Gunawardena, in press at @eLife.
In the process, we try to resurrect the neglected contributions of Beatrice Gelber.
4
37
176
What's interesting to me is that we demand a level of explainability from AI that we can't even satisfy when we explain our own behavior, let alone other people's behavior.
15
21
180
@NatureNeuro please consider eliminating the reference limit. It contributes to a rich-gets-richer pattern where only a small number of papers get cited over and over, and these tend to be by men.
4
11
155
Surely the greatest indictment of reinforcement learning theory is the fact that young children cannot be trained to control the volume of their voices no matter how much reinforcement.
6
10
168
This is a fascinating story:
quantamagazine.org/insect-br…
What perplexes me, though, is the title: the memories DO survive!! What doesn't survive is the anatomy.
5
30
165
22,865
I'm concerned that a disproportionate amount of reviewing load is falling on a subset of eligible reviewers. I noticed someone I was going to invite to review hasn't accepted an invitation for the journal since 2012, despite publishing there repeatedly during that time interval!
14
7
154
Can we stop saying that tasks have been "solved" by AI systems? This sounds like there is nothing left to do, when usually there is a lot left to do. For example, attaining human-level performance asymptotically is very different from attaining human-level sample complexity.
5
12
167
Today my laces got wrapped multiple times around my bike pedal. As I hurtled down a hill, I imagined the headline: “Harvard professor dies at the hands of his feet.”
10
2
155
33,394
I'm a priori sympathetic to Haidt's position (and I will still restrict social media for my kids), but I find his cavalier attitude towards evidence incredibly frustrating. It would be better if he just said these were his opinions.
4
12
154
33,960
I don't understand why CS conferences rely on a peer review system in which real revision is not possible. At @NeurIPSConf for example, author responses are limited to 1 page and there is no follow up; reviewers never get to see a revised manuscript.
7
15
167
@LanguageMIT and I are so pleased about this:
mitpress.mit.edu/cognitive-s…
Thanks to @HarvardLibrary @mitlibraries @mitpress for making this happen!
3
40
151
16,641
But why are we doing brain-wide association studies in the first place? I think the real implication is that we should be thinking more deeply about theory in order to guide experimental research. Successful science is not a dragnet.
Similar to #GWAS, we think this work suggests that larger datasets and consortia (N⪆2,000) are needed to usher in a new era of reproducible human brain-wide association studies - #BWAS. Of course, however, smaller studies will continue to be pivotal in the field #MSC.
3
24
149
PI just had an idea and no one wants to say it's dumb.
“What kind of lab are you running here? nitter.app/happyyouare/status/133…
2
9
145
Did you know that Apollinaire, Klimt, and Schiele all had some of their most unproductive months during the Spanish Flu?
7
8
142
There should be an incentive system for reviewing where you receive credits for publication fees when you review for a journal. So people who submit more than they review should pay a larger share. Could also incentivize on-time reviews in this way.
8
7
132
The paradox of hopeless optimism, from Francis Crick. I think this condition also afflicts experimental neurobiology. Many of us persist in using methods that are unlikely to yield answers to our questions. This persistence is aided by the imprecision of our questions.
4
21
157
We can accelerate progress in AI by providing free daycare to all grad students and postdocs working on AI. Ask your funding agency to support daycare-inspired AI!
3
11
146
Everything I've seen about the metaverse is so boring. I want to experience what it's like to be in another person's body, or to see the world through a fly's compound eye, or to taste a soup cooked 8000 miles away, or to deeply understand the proof of Fermat's last theorem.
20
9
144
@profcikara, who has extravagantly sent 👑👑👑👑 to countless other individuals, is now herself 👑👑👑👑👑👑👑👑👑👑👑, having attained the status of tenured boss-level professor-queen!
I couldn't be happier to have her as a friend and colleague forever.
7
4
136
18,701
Over a decade ago, Josh Tenenbaum gave a memorable talk at Cosyne in which he argued that the language of neuroscience was essentially the language of 70s EE, and contrasted it with the modern language of CS. It's interesting that not much has changed since then. Why?
11
12
146
What is the strongest argument for "emergence" in complex systems that doesn't boil down to a form of simplicity (coarse-graining is easier to think about / work with)?
25
18
141
34,089
I'm hiring a full time research assistant / lab manager! sjobs.brassring.com/TGnewUI/…
7
89
137
What is a paper that you loved *after* doing a journal club on it? I find that journal clubs often end up being overly focused on the flaws of papers.
13
2
134
33,778
Provocative correspondence in Nature today:
nature.com/articles/d41586-0…
I have to admit that I don't really understand these arguments, though...
3
26
133
35,365
This is great. I would love to teach a course on "The art of spontaneity". No syllabus. Each class I'll take students for a walk and confront them with something or somebody totally unexpected. The goal is bewilderment punctuated by flashes of insight.
Fantastic piece by Marten Scheffer on the forgotten half of thinking:
".. taking walks, reading things unrelated to your research, and hanging out with strangers in a campus pub should be considered part of the serious process of thinking .."
pnas.org/doi/full/10.1073/pn…
11
6
122
20,518
1550 BC: I think it's clear by this point that blood-letting is going to be the only relevant technique in the future. There is something magical about it. Against all statistical convention its performance improves with the amount of blood let.
I think it's clear by this point that NNs are going to be the only relevant technique in the future. There is something magical about them. Against all statistical convention their performance improves with the model size. Free up time and teach important ideas form other fields!
1
11
129
Can someone explain to me how this is different from the well-known result that doing linear PCA on place cells gives you grid cells? You can get this without the LSTM component.
3
25
129
This is (very) old news, but apparently the action potential was first discovered in the Venus fly trap!
royalsocietypublishing.org/d…
One spike to open the trap, one spike to close it. These experiments were done based on speculations by none other than Charles Darwin.
1
15
118
Can we agree to not use the words "dopaminergic midbrain" when talking about fMRI results? There are other kinds of neurons in the midbrain and fMRI does not measure dopamine release.
5
6
122
I'm hiring an archaeologist to help figure out scientific data and code downloaded from repositories. Applicants should be skilled at deciphering unannotated and disorganized artifacts deposited without regard for comprehension by future civilizations.
2
9
117
12,001
I'm curious what would happen if we banned possessives when referring to hypotheses in scientific papers. My impression is that scientists are far too personally invested in their hypotheses. When I edit my student's papers, I change 'our hypothesis' to 'a hypothesis'.
15
6
124
15,103
3yo has been demanding that my wife tell him spooky stories while he sits on the potty. She's literally scaring the shit out of him.
1
119
I'm looking to hire a computational neuroscience postdoc to work on models of the basal ganglia and dopamine. More info here: gershmanlab.webfactional.com…
@u19teamdope @CogNeuroJobs
62
115
I completely agree that grants should be works of art, which is why all my grants shred themselves immediately after being reviewed, Banksy-style.
To all the grant writers out there: your grant not only has to be scientifically innovative, significant, rigorous etc--but it also has to be a well-written work of art that moves the reviewers to proselytize your work. Beautiful illustrations are as important as the prose.
2
6
102