Delighted to share our
#neurips2023
paper w
@grockious
@hmd_palangi
et al
Evaluating Cognitive Maps & Planning in LLMs with CogEval
We test planning in 8 LLMs.
Failures like hallucinating invalid paths/falling in loops don't support emergent planning.
1/n
Honored & excited to start as a Senior Researcher in Reinforcement Learning
@MSFTResearch
NYC!
I'll continue to build models & run experiments alongside the incredible RL team-& AI fairness & comp social sci teams-whom I've admired for long.
Grateful for supportive mentors 🙏🏼 1/5
Excited to say I was promoted to Principal Researcher
@MSFTResearch
NYC. Cousin wondered if now I chase researchers in the yard & call ppl to my office but
@hannawallach
says it’s about intellectual innovations with creativity & freedom, & being the goto in my field.
Grateful😊
After 15 years, & having done neuroscience & computational research during this time, I am rereading this anthology of philosophy of mind papers. Organized & each paper only few pages. If anyone else is interested, we could communicate reading via tweets or smoke signals or such.
Delighted that my paper
"Collective Minds:
Social Network Topology Shapes Collective Cognition"
is out in Philosophical Transactions of
@royalsociety
issue:
‘The emergence of collective knowledge and cumulative culture in animals, humans & machines’
Honored that my paper is out 🤓 alongside amazing papers 🧠 in an issue by the wonderful
@morganbarense
&
@hugospiers
🙏🏼!
I review learning structures w the reinforcement learning framework,
& relation to Abstraction, Intrinsic Motivation, & Exploration.
New! "Learning Structures: Predictive Representations, Replay, and Generalization"
Flexible memory & planning rely on learning compact structures of states. I review how reinforcement learning can acquire structures in absence of rewards.
for
#cosyne2020
After 14 months indoors I spent 3 days hiking in
#Yosemite
& 2 by the ocean. Can't adequately express the healing, the phenomenology, or the perspective the mountains afford. Every emotion rises on a 6-mile hike & you reconnect to self/nature on the way down.
I was born for this.
My dad finally got his first vaccine shot today in Tehran!💉
I’m just back from 2 weeks of visiting him & my sister after 5.5 years of separation by bans &pandemic.
Our family love language is discussing theories, mythologies, philosophy, science fiction at the kitchen counter❤️.
Is this a joke? $12,000 to make a paper open access that we submit & review for free?? Even in industry research I can’t justify such a cost.
It’s the price of
-1 year rent for room in NYC
-5 cars
-10 laptops
-1/3rd annual PhD student salary
-1/2 annual nyc min wage
Outrageous.
In this scenario, the cost of publication is covered by an Article Processing Charge (APC) paid at the time of publication. The APC for Nature Neuroscience in 2022 is €9,500/US $11,390/£8,290.
New! "Learning Structures: Predictive Representations, Replay, and Generalization"
Flexible memory & planning rely on learning compact structures of states. I review how reinforcement learning can acquire structures in absence of rewards.
for
#cosyne2020
"Mirzakhani did poorly in math in middle school. But a change from [a teacher] who didn’t see her talent to one who did, made all the difference.
“It’s so important what others see in you”
In 2014 she was the first woman & Iranian to win the Fields Medal"
I've studied memory for 15 years with an empirical lens, scanned it, modeled it.
"The logic of memory" is my philosophical musing on memory, what I found missing in scientific discourse.
Thanks to
@Philosoph_Salon
&
@esdsantos
's course on James Baldwin🙏🏼
🧠Exciting new work w
@TaylorWWebb
&
@ShankaMondal
@MSFTResearch
A Prefrontal Cortex-inspired Architecture for Planning in LLMS
LLMs struggle w planning.
Our solution is inspired by brains: planning via recurrent interactions of multiple PFC subregions.
🧵
Hi! So excited to share our newest w
@IvaBrunec
:
Predictive Representations in Hippocampal & Prefrontal Hierarchies.
fMRI evidence that predictive representations are organized at multiple scales in posterior-anterior hierarchies in PFC & hippocampus. 1/N
Hi, thanks to Lynn Nadel & Sarah Aronowitz I wrote a book chapter on multiscale predictive representations that connect the past to the future, memory to planning.
There's still time to edit it & I'd love to read your comments!
Am I missing citations?
@elonmusk
@Neuro_Skeptic
You do realize many of us record single neuron data in humans, rodents, monkeys, etc right? That neural recording AND stimulation have been going on for over 60 years? Of course someone must have told you, where else would you get the idea.
Dear colleagues with Iranian trainees in neuro/psych/cog/AI,
Holding mental health zoom check-ins for our turbulent times.
To ensure a *safe space*: Iranians I know & Iranian women with a trusted reference are welcome to DM.🙏🏼
@yasser_roudi
@AthenaAkrami
@ShahabBakht
@theHessam
Excited to share SAPIENS!
W
@nisioti_eleni
@pyoudeyer
@Clement_MF_
we had multi-agent deep RL solve hierarchical innovation problems & found:
The Social Network Topology of Experience Sharing among Agents Shapes Innovation
🧵1/n
Today's the 6th anniversary of my mom's passing.
Lida was a political scientist, NGO leader, & litigator representing workers. She identified as a social democrat & enjoyed singing, hosting dinners, & reading fiction & Jung's work.
Still mourning. But grateful her light is w me🕯️
New paper alert 📢📢
A Rubric for Human-like Agents & NeuroAI
Cog, neuro, & comp sciences increasingly reference human-like agents & neuroAI w no consensus on scope/use. Some import progress in 1 discipline another &assume it automatically translates.
1/n
I've found Microsoft Research to be more academic & creative than I had anticipated with a shockingly flat hierarchy. MSR is hiring at all levels including
paid research interns (summer 2021), postdocs (2-year), RL researchers, senior researchers, & principal researchers.
Apply!
Many open Reinforcement Learning related positions are here: . Please consider applying if you have RL-complimentary expertise and want to discover and build the future.
Friends of the
#learningSalon
, please mark your calendars & pen your book margins as we host:
György Buzsáki,
"The brain from the inside out"
Friday Jan 28, 4 PM ET.
Event link:
The book:
Buzsaki lab:
I’d been surprised by the stark difference in reaction to Ukraine vs Palestine or Afghanistan. Cue the dehumanization of Afghanistan & Iraq, among the oldest civilizations, saying “oh no not in civilized Europe where we don’t expect this”. Not where the US/UK are used to bombing.
"This isn't a place, with all due respect, like Iraq or Afghanistan who has seen conflict rage for decades. This is a relatively civilized, relatively European - I have to choose those words carefully - city where you wouldn't expect that or hope that it was going to happen."
I know everyone raises this every year, but why isn't there 1 portal to submit 1 reference letter per grad student on the job market that any university can access instead of having to submit to 500 different links?
I'm excited to teach python to formerly incarcerated individuals at
@Columbia
's Justice Through Code program. (Center for Justice:
@JusticeColumbia
)
⚖️💻
Class completion awards students certificates to increase probabilities of future employment.
Can we have academic accountability groups? Everyone is making promises about improving diversity but as Angela Davis put it we know this intense moment will be over, what happens after? Who will check in with labs& departments& journals in a year to assess their actual progress?
Excited to share new work w
@katjahofmann
@smdvln
@ralgeorgescu
@JarekRzepecki
Evelyn Zuniga & colleagues!
Navigation Turing Test (NTT): Learning to Evaluate Human-Like Navigation
accepted
@ICML
We propose a method to evaluate human-like navigation
🧵1/n
Not sure about twitter's future, but gotta say how grateful I am for science twitter.
You have been my community through the darkest times of the pandemic, as I switched jobs, as we launched the salon, during Iran's revolution.
I learn from you every day, even from our arguments.
Science twitter please don’t leave. We have great community here. I know no other place to keep up and interact with ideas from neuroscience, psychology, physics, mathematics, machine learning, philosophy, economics & linguistics. No reason to stop; just block others if needed
📢 📢 Delighted to share our paper with the amazing team Safoora Yousefi, Leo Betthauser,
@grockious
@raphaelmilliere
on neuro-inspired analysis of LLMs.
Decoding In-Context Learning: Neuroscience-inspired Analysis of Representations in LLMs
🧵1/n
#Llama
Relieved to finally receive travel and work permit via my greencard application! I still need a visa for everywhere else, but will be able to work & see my family after years!! Huge relief... thank you to scientists & friends who wrote my letters & supported my application! 🙏🏼😊
We need more papers like this on what ML can learn from Cog/Neuro. Adding PFC can advance ML capacity for Reasoning, Compositionality/systematicity, Abstract Planning, Control in novel env.
DEEP LEARNING NEEDS A PREFRONTAL CORTEX
Russin, O'Reilly, Bengio.
A year ago I didn't understand how people lost track of emails. I sometimes replied a bit late or missed an email or 2 but always had a mental grasp of emails: people, reviews, talks, who needed a reply.
Now I'm lost.
I constantly feel email debt.
Do ppl declare email bankruptcy?
@KordingLab
Konrad, he also said COVID will be over in April. Generating random statements, without adequate knowledge but with unearned confidence to serve one's business, is NOT reasonably defendable-even if Konrad agrees w it. Classic non-expert overconfident techbro elevating themselves.
I moved to industry research June 2020. Worked hard to keep & strengthen my comp neuro research & coauthorships, Mentoring, reviewing, talks etc.
At a Neuro AI conf I realized it paid off, thx to the approval of my community, discussions, & laughter w my coauthors.
I Missed this.
You send a tex file with all the correct equations to a journal.
They say no send the MS word file, then they make a million errors in the equations.
You spend weeks back&forth correcting proofs.
Why? Please use overleaf. It's free & will save you and your authors a lot of time.
I miss being in academia.
For the first time since I’ve joined MSR (2020) we have all our grad student interns around in person this summer. The learning & research energy fills the space & it’s the most academic our offices have felt.
Realized I missed this.
I miss academia.
📢Are you (or do you know) a PhD student with experience in transformers, LLMs, & cognitive neuroscience (esp PFC, hippocampus)?
Still interviewing interns for a project following up on our earlier work:
job:
Tested positive for covid after a conference 😞 in spite of being masked, measuring CO2, & begging everyone to wear masks, I let my guard down at the final party.
Please send advice, medication recs, pet pics, research papers to help avoid long covid 🙏🏼
P.s.
@AthenaAkrami
👀
I used to think of time as an arrow.
I was moving into the future.
Right now time feels like guitar hero:
A set of calendar events that come at me at rapid speed with varying difficulty level & overlap,
& I need to adapt fast to play songs I've never heard.
This was my TED talk.
Teaching tensors of different ranks and basis sets take time, you always need to use visual aid to teach how to imagine it. I usually draw book shelves to visualize it. This video does a better 3D visualization job, especially the last 3.5 minutes.
Looking for papers & refs using manifolds & geometry for analysis of large scale behavior. It’s common to use it in computational neuro, but wondering about refs focused on behavior (even if it accompanies neuro) 🙏🏼
Example refs:
@SussilloDavid
@FieteGroup
Tank lab
@StefanoFusi2
Nov 29-Dec 19. I’ve been farming salad under artificial light in my kitchen. Mr. Robot’s been watching over - & drowning in - them as you can see. It’s magical, endearing, almost alien to observe closely their daily growth, the way they curve, the colors. How could I eat them!?😱
Delighted to announce that
@katecrawford
& I are hiring a MSR-NYC intern for a project on the intellectual history of machine learning.
If you are a *PhD student* in History and Philosophy of Science, STS, or related field please apply!
*Deadline April 4*
Excited to hear our paper got accepted in
@eLife
(more will follow).
I'm impressed: eLife gives you a DOI & publishes your work *within DAYS after acceptance*!!! Plus eLife is open access *and* you can see the review process.
This is such a good model.
Visa approved after 2 applications=> Excited to finally announce: I'll be working w
@neurojosh
at Columbia University for a year. We design VR revaluation tasks for *Human Single Unit Recording* to test predictive map models @ neuron level!
#dreamproject
#successorRepresentation
Is there experimental work on the psychology of human bonding with plants?
Is there a culture/religion in which taking care of plants is a ritual or moral imperative? Iranian ritual: grow greens each new year (spring).
Ps. I ❤️ my plants! & my monstera has grown ~4x this summer.
It's the 4th anniversary of my mom's passing. Lida, super mom, political scientist who saw work a replenishing source of pride &energy, lives within me. Half my DNA is hers. She lives in my psyche, my desires, determination, resilience & optimism. I celebrate her life & legacy ♥️
Friends, an internship, AI residency, & multiple postdoc positions are open at Microsoft Research Cambridge, UK, on:
Deep Reinforcement Learning for Games
You get to work w brilliant scientists I've had the pleasure of working with!
@katjahofmann
@smdvln
Hi science twitter! I'm on the job market, & though I constantly check, I surely will miss some job ads! If you know of an assistant professor position related to cognitive computational neuroscience (neuroscience/psych departments, comp sci, cog sci, related) please DM me!🙏🏼
Many neuroscientists assume that reinforcement is just model-free RL, the same as conditioning, or that it needs external rewards. I have meta-learned to clarify that is not the case in early slides of my talks. Sergey Levine's slide below is great at making this crystal clear.
They re"discovered" citation bias that
@DaniSBassett
lab had identified: Men cite W-W papers less often.Millennia of sexism➡️more men + misogyny in academia. Undercitation adds up.
UAE authors attack diversity by making men's underciting the measure of women's failure! Amateurs🤦🏻♀️
@david_m_amodio
Please note this is not about all fMRI research but v short 8min resting state functional connectivity/RSFC & cortical thickness for clinical correlations-w much existing critique.
Most task-based fMRI studies use 40-80min fMRI w carefully designed paradigms w specific hypotheses
The
#LearningSalon
began Sep 2020! Grateful for having a community to discuss contentions & bridges in biological & artificial learning🧠🦾 every Friday. Multidisciplinary year-round discussion w respectful disagreement is a neglected skill. We'll continue practicing it in 2021🙏🏼
Excited to share new work w Yi Wan, Ali Rahimi, Jana Rajendran
@harmvanseijen
@apsarathchandar
showing popular deep model-based methods (PlaNet, DreamerV2, MuZero) perform poorly at adapting to local environmental changes.
We explore causes & solutions
1/n
Happy 2020🎈 Been a decade of loss & growth. 2 surgeries+PT during postdoc. Lost 2 women who raised me: grandma & mom. Trump/travel ban. Faced PTSD. Lost friends to a shooting.
Got PhD. 17 papers &chptrs. art collabs. New friends❤️ Healthy!
Wishing more health & starting my lab🍾
I can't wait till the Learning Salon starts up again!
In the meantime, a paper on lessons learned, from the organizers,
@criticalneuro
@neuro_data
and
@blamlab
There is such a satisfaction in finding out and fixing a small bug and making thousands of lines of code work and get the results. You go from frustration to victory in a matter of hours. There should be a word for it, a haiku, even a scent. I'd wear that perfume all the time.
Our newest w
@alinicoman
,
@AjuaDuker
!
Bridge ties bind collective memories
In clustered networks, timing of convos matter. If those bridging clusters talk to each other first, memories converge more. Too late to converge if all talk w their bubble first.
Excited to say I'm hiring a PhD student for a 12-week internship w myself &
@harmvanseijen
to work on task-sets.
Get in touch if you:
- have strong RL background
- are experienced with transformers
- will still be a PhD student by internship day 1 (~June)
Thrilled to share new work w Stacey Sinclair &
@profcikara
! Computational Justice: Simulating Structural Bias and Interventions.
We ran agent based simulation of structural bias, params set from studies.
Then simulated/compared different interventions.1/n
Bat signal to academic and industry scientist friends:
If you are writing or know of a petition against the visa ban order please DM me. A few of us hope to centralize efforts.
✊🏼🙏🏼
In recent weeks many wildly successful & prolific colleagues were laid off from google, Microsoft, meta. A triggering spectacle. 7 years ago my mom was laid off, had a stroke the next day, & passed away after 50 days in coma. If you need me I’ll be meditating on that among trees.
Excited to share our new preprint with Marc Howard, "Predicting the future with multi-scale successor representations". Mathy, but we've tried to give intuitive explanations of equations. Would love to hear thoughts!
Excited to give a talk (in an hour) to a large in Person audience at the
@ChampalimaudF
in Lisbon (first in person conference since the pandemic). Tough act to follow Peter Dayan & joyful to have
@neuro_kim
&
@blamlab
in person in the audience :)
#csr2021
International friends: the US Supreme Court lifted New York’s ban on guns in public & overturned the right of abortion, criminalizing it. Both will have global consequences too.
If you find us irritable these days, many of us, not just women & NYers, are in a state of mourning.
Thank you for this brilliant suggestion! I started reading "Other minds" 2 days ago and cannot put it down! Fantastic science writing, thought provoking, original and critical thinking by
@pgodfreysmith
.
@criticalneuro
Maybe not exactly what you're looking for, but there's some really interesting discussion on this in Chapter 7 of Other Minds by
@pgodfreysmith
!
@EricMTrautmann
Wireless recording existed. They scaled it. great incremental engineering NOT as revolutionary as claimed. Idk what you mean by negativity but experts in the field are right to critique over the top claims/nonsense abt consciousness.
Why the negativity & defense against experts?
At the
#LearningSalon
(lightly paraphrasing)
"Make your opinions extreme enough so they can be attacked. If you don't attack my opinion and I don't attack yours, we don't understand each other."
- Buzsaki
Friends of the
#learningSalon
, please mark your calendars & pen your book margins as we host:
György Buzsáki,
"The brain from the inside out"
Friday Jan 28, 4 PM ET.
Event link:
The book:
Buzsaki lab:
Some see the move btwn academic & industry research as binary. During the pandemic I've kept very close to academia thanks to the Learning Salon, zoom talks @ departments /confrnces, mentoring students.
But I truly miss working @ a university campus!
Realizing my mind never left.
Join us today Friday 4 PM ET to hear the amazing
@jhamrick
. Jess's work is a fantastic cog sci+AI hybrid. She gives some of the best tutorials/talks at ML conferences, sometimes wearing a construction hat.
I recommend her review on DL & mental simulation:
The Learning Salon hosted by myself &
@blamlab
is super excited to host
@spiantado
tomorrow, Friday 4 PM ET, to talk about:
"Church Encoding as the link between Cognition and Neuroscience"
Related reading:
Join us
#learningsalon
!
It was fun to be interviewed by
@pgmid
for the Braininspired podcast. Paul did a great job of making me share & bridge life stories & research in ways I hadn’t experienced. I appreciate connections drawn by convo & the chance to discuss the bigger picture,RL, neuro, Philosophy.🙏🏼
Ida
@criticalneuro
and I discuss why she's always cherished learning myriad subjects, the role of philosophy in her world, the reinforcement learning landscape, successor representations in RL, and how to name constellations.
Many science fiction stories mention uploading consciousness into a machine. Do you know of stories where they explain how they do this? No matter how banal or scientifically crazy, I’d love to read about those references if you recall any.
@_Quazi_
@anorangeduck
(3/5) After a day of training, he can withstand significant perturbations (but sometimes you just get too many cubes in the face…):
Excited to share our new work w
@beausievers
!
When we remember something, its associates are a bit activated too. Our model, SAMPL: Spearing Activation and Memory PLasticity, models both forgetting & enhancement of those further associations in memory. 1/n
Thanks for this, been weighing on me but didn’t share: Twitter is less supportive lately & the struggle against imposed hijab in iran is not to be confused w western islamophobia or a cry for neocolonial intervention.
Iranian women:
drop a DM if I can offer support💚
Love & Power
A 22 year old woman named
#Mahsa_Amini
has been beaten to death by the Iranian “morality police”. As I’m writing this many have been beaten and some have died in the protests that broke out after. You might ask what the hell is a morality police and why is this happening? …
What do you want to see in computational models of replay that is missing from existing models? I'd especially love to hear from biologists & electrophysiologists –recently I gave a talk at the Buzsaki lab & I'm still inspired by their questions.
Tweet/respond/DM/smoke signal.🙏🏽
📢 In 2 days, Thurs Jan 14th, we'll host RL day!!
Come for the debate on theoretical RL vs deep RL between
@JohnCLangford
of
@MSFTResearch
vs Yoshua Bengio of
@Mila_Quebec
moderator:
@katjahofmann
,
or Doina Precup's talk moderator: yours truly,
& stay for student posters!
🤖🦾⏳👇
There’s still time to register for this year's Reinforcement Learning Day on January 14, 2021. Hear from RL experts like Professor Doina Precup
@mcgillu
, Professor Yoshua Bengio
@Mila_Quebec
, and Dr. John Langford
@MSFTResearch
.
Reserve your virtual seat:
Next meta will propose a novel alternative vision for time, which will call the temporal window during which a location of earth receives sunlight (even if cloudy) “day” and the rest “night”. They’ll add a sketch too, which is a collage of textbook images from 3 decades ago. 🤦🏻♀️
Chief AI Scientist Yann LeCun (
@ylecun
) is sketching an alternate vision for building human-level AI. LeCun proposes that the ability to learn “world models” — internal models of how the world works — may be the key. Learn more: