1/ I'm excited to share our new paper, now in
@Nature
!
Paper:
PDF:
Summary 🧵:
We present a new theory of problem simplification to answer an old question in cognitive science and AI:
How do we represent problems when planning?
I’m excited to share some news: My group and I are moving to
@NYUPsych
!
I’ll be affiliated with both cognition/perception and social psych, and I am SO thrilled about this unique opportunity to help bridge the two programs 🤗
Excited to share that Fall 2023, I'm starting a lab in the CS department
@FollowStevens
!
Even MORE excited to share that I'm recruiting Ph.D. students interested in computational Cog Sci, RL, and/or HCI 🧠💭🤖!
The official deadline is *Feb 1st*...
Excited to share a new paper just accepted at Psych Science 🎉
Preprint:
We look at one of my all time favorite cognitive biases:
✨💭 Functional Fixedness 💭✨
Have you heard the idiom “To a person with a hammer, everything looks like a nail”?
A 🧵
Incoming Faculty Fellow Mark Ho (
@Mark_Ho_
) currently works as a post-doc in the Computer Science and Psychology departments at Princeton University and will be joining CDS this fall. Read more about Mark on our blog!
#datascience
Now in press at Annual Review of Control, Robotics, and Autonomous Systems! Tom Griffiths (
@cocosci_lab
) and I take a whirlwind tour of research on human decision-making and theory of mind in relation to control and robotics 🤔 💭 🤖 1/5
New preprint up! 📜
"Bayesian Reinforcement Learning with Limited Cognitive Load" with
@Dilip_Arumugam
@noahdgoodman
and Ben Van Roy
This was such a fun and stimulating paper to work on 🤓 A few additional points to complement Dilip's excellent summary:
Q: What happens when you combine Bayesian inference, reinforcement learning, and rate-distortion theory?
A: A way to formalize capacity-limited cognition in biological and artificial decision-making agents!
New paper with
@mark_ho_
,
@noahdgoodman
, & BVR -
SUPER excited to share a new pre-print on how people break down big problems into smaller, more manageable ones 🎉🎉🎉
This work was done with
@cocosci_lab
@nathanieldaw
@callfredaway
and spearheaded by the brilliant
@_cgcorrea
🙌🥳
✨Preprint✨ w/
@mark_ho_
,
@callfredaway
,
@nathanieldaw
, & Tom Griffiths:
TLDR: How do people decompose goals into subgoals? We ran a large-scale experiment to test how the computational cost of planning drives subgoal choice.
More in 🧵
By the time a dog, cat, or chimpanzee turns 5, its seen a similar number of "frames" to learn how the world works through self-supervised learning.
But a 5yo human can do things a 5yo dog/cat/chimp can't (e.g., use language productively, etc).
By the time a human child turns 5, she has seen the equivalent of 800 million "frames" of video + audio + touch to learn how the world works through Self-Supervised Learning.
Much of it is acquired actively.
5 years * 365 days * 12 hours * 3600 seconds * 10 fps = 788.4 million
So excited for
#RLDM2022
and Saturday's
#RLasAgency
workshop, organized with the amazing
@aharutyu
+
@dabelcs
🤩🤩
We'll be discussing the limits+possibilities of RL as a model of agency in cog sci/neuro/philosphy/AI with a brilliant group of scholars...
At
#AAAI2020
?
I'll be talking today about our paper, "The Efficiency of Human Cognition Reflects Planned Information Processing" w
@dabelcs
, Jon Cohen,
@mlittmancs
, and Tom Griffiths.
Talk at 2pm in Madison
Poster
#44
at 6:30-8:30
Link to paper below!
Excited that our paper "People teach with rewards and punishments as communication, not reinforcements" is now out in JEP:General 😃 preprint here w/
@JoeAusterweil
@fierycushman
@mlittmancs
Caught a few utils last week visiting the auto-icon (mummified body of Jeremy Bentham) at the UCL Student Center
I have my complaints about utilitarianism, but I can respect the commitment to the ideas!
(He thought widespread voluntary mummification would maximize utility)
Excited to be in SF for jam-packed
@CogCompNeuro
2022!
First, I’ll be presenting recent work w Jon Cohen and
@cocosci_lab
on *rigidity in planning* 🤔💭
Poster P1.24: Tonight at 7:30
Talk: Tomorrow at the 10:30-12 session
New preprint w
@JoeAusterweil
@fierycushman
@mlittmancs
on communicative demonstrations! We model how people communicate with their intentional actions as well as report new experiments + reanalyze prior developmental work. Reach out if you have feedback!
Our paper on what makes abstractions useful is out!
We provide a bird's eye view of how abstract representations help solve hard RL problems and what this may tell us about people. Really enjoyed working on this w
@dabelcs
T Griffiths and
@mlittmancs
!
Had such an amazing time at Diverse Intelligences Summer Institute last week, catching up with old friends and making new ones!!
Until next time!🦋👶🐵🤖👾👽🏴
#disi2023
#standrews
#scotland
New paper in Cognition led by the amazing
@tedsumers
!!
We examined how people teach using demonstrations ✋versus language 🗣, and under what conditions each strategy works better (eg the complexity of what’s being taught). Check it out!
Humans use a wide variety of teaching strategies. To teach you chess, I could *demonstrate* how rooks move (dragging one across the board) or use *language* to explain ("this piece moves in straight lines"). Our new work in
#cognition
compares modalities!
Daniel Kahneman at last week’s
@EffortMental
workshop:
“I don’t think people are rational but I do think they’re reasonable, and rational can be a good approximation for reasonable”
As the old saying goes: instruct someone to fish, feed them for a day; describe to them how to catch fish, feed them for a lifetime (or something like that? 😋🎣)
Anyway check out our new pre-print on teaching with *descriptions* vs *instructions* in RL, led by
@tedsumers
🎉
What kinds of language afford generalization for RL agents? Our new preprint presents a formal framework for inferring humans' reward functions from utterances. We show that *descriptions* generalize better than *instructions*.
5/ Our paper develops a formal theory of this intuitive notion of value-guided construal.
The key idea is that planning involves not only selectively simulating *actions* (as is standardly assumed), but also selectively representing *problems* (i.e. forming task construals)
So excited for
#RLDM2022
and Saturday's
#RLasAgency
workshop, organized with the amazing
@aharutyu
+
@dabelcs
🤩🤩
We'll be discussing the limits+possibilities of RL as a model of agency in cog sci/neuro/philosphy/AI with a brilliant group of scholars...
I noticed this morning that the cover of “Artificial Intelligence: A Modern Approach” (3rd ed) by Russell & Norvig includes... a picture of someone reading “Artificial Intelligence: A Modern Approach” (3rd ed) by Russell & Norvig. 🤔
#recursion
Dear cat cognition researchers,
Are cats generally able to do the kind of planning / look ahead / physical reasoning suggested in this vid? I’m genuinely curious!
New paper with
@vaelgates
,
@callfredaway
, and
@cocosci_lab
! We examine how people learn about other people's thoughts and preferences based on *how long it takes them to make a decision* ⏲️
If you ask someone out for coffee, which would you rather hear?
Sure!
or
... sure.
The answer is the same, but their *response time* tells you something about how much they like you. Our paper in Cognition formalizes and tests this idea. 1/6
My first preprint! How do people use rewards+punishments to influence others’ behavior? They tend not to “shape” (in the classic RL sense often assumed) and instead use r+p to communicatively *signal* value. w
@JoeAusterweil
@fierycushman
@mlittmancs
Excited to share a new paper just accepted at Psych Science 🎉
Preprint:
We look at one of my all time favorite cognitive biases:
✨💭 Functional Fixedness 💭✨
Have you heard the idiom “To a person with a hammer, everything looks like a nail”?
A 🧵
Really excited about new work with
@_cgcorrea
,
@callfredaway
, and Tom Griffiths that models hierarchical planning and subgoaling as a resource-rational representation problem. Come check out our poster at
#CogSci2020
!
@Mark_Ho_
,
@callfredaway
, Tom Griffiths and I will present work at
#CogSci2020
that models how—and *why*—people decompose big scary tasks 🐲 into smaller, more friendly tasks that make it easier to plan 🧩🐭🐭🐭.
Poster P-3-566, Zoom on Aug 1. Preprint:
✨New preprint✨
In this work led by the amazing
@_cgcorrea
, we examined the hierarchical structure of human problem solving using a novel programming task! 🤔🤖
Human behavior is hierarchically structured. But what determines *which* hierarchies people use? In a preprint, we run an experiment where people create programs that correspond to hierarchies, finding that people prefer structures with more reuse.
1/7
4/ As an answer, we propose that when people plan, they actively construct simplified, but useful, models of the problems they face—much like how maps are simplified but useful models of the world.
We call this process of cognitive map-making *value-guided construal*
2/
50+ years ago, Allen Newell and Herbert Simon introduced the idea that planning consists of two parts:
1. Representing a problem (e.g. modeling the pieces and rules of chess), and
2. Simulating actions within that representation (e.g. doing heuristic search using a model)
We're excited to host
@mark_ho_
for Wednesday's Consciousness Club on "Construction of mental representations in human planning"
Wed 8th Feb 3pm-430pm GMT
For more information and how to join, please see
All welcome!
@LeverhulmeTrust
@EP_UCL
@WCHN_UCL
This is a fantastic example of changing one’s task construal:
By refocusing your attention to different affordances in a situation, the right solution becomes immediately obvious!
Cognitive scientists! You don’t want to miss our
#CogSci2023
Workshop: “How does the mind discover useful abstractions?” Co-organized w/ the peerless
@wkvong
+ Lio Wong +
@marcelomattar
!
Currently redoing my academic website and finding
@Alpine_JS
to be a very nice lightweight way to write responsive components without the overhead of react/vue/etc
What makes something (or someone) an agent? 🤔
Is reinforcement learning a useful way to conceptualize agency? 🧐
If you find such questions interesting, consider submitting to the
#RLDM2022
workshop that
@dabelcs
,
@aharutyu
and I are organizing! 😁
🎊 Excited to announce the “RL as a Model of Agency” workshop at RLDM @ Brown! 🎉
We invite submission of 1-3 page papers that draw attention to a question or perspective related to the question "Is RL a suitable model of agency?"
Full details:
I'm excited to share that my paper on resource-rational task decomposition has been published in PLOS Computational Biology! 🎉🥳
TLDR: People break large tasks down into smaller ones by balancing efficiency and the computational cost of planning.
I was invited by
@sciencebreaker
to write a "break", a short summary for non-academics, about our work on value-guided construals. Excited to see it out!
✈️🤔 Ever been stuck due to a canceled flight and had to re-plan your journey? A new study delves into "value-guided construal" – our ability to plan flexibly between utility and complexity.
#scicomm
#openscience
#thesciencebreaker
Let's break it down:
Had a great time exploring the universe of social and collective behavior at
#COSMOSKonstanz
2022 🚀🛸
Thanks
@TheCharleyWu
and
@WataruToyokawa
for bringing together an amazing group of people and introducing us the beautiful city of Konstanz 🤩!!
✨ Now out in Psychological Review! ✨
We present a new account of relevance that weighs both **epistemic** and **decision-theoretic** utility: statements are relevant if they improve the listener’s future decision-making.
What happens when you put your foot 🦶 in your mouth 😮? 🤔
Check out our
#cogsci2019
preprint on unintentional speech acts! We model the role of false assumptions about common ground in
#awkward
situations (eg faux pas).
w J Korman & T Griffiths
I recently had the pleasure of writing about some of our work on construal and functional fixedness for the
@SPSPnews
blog Character and Context - you can check it out here!
"As our research shows, these errors follow from the reasonable desire to simplify the world and the ease of sticking with an initial simplification. " -
@mark_ho_
If you're at
#ccn2018
, come hear us talk at 4:30 about how people interpret others' problem solving behavior using hierarchical representations! (Poster Session 1A
#38
) Joint work with
@sophia_sanborn
, Fred Callaway, David Bourgin, & Tom Griffiths
What a way to start an abstract:
“Does gravity affect decision-making? This question comes into sharp focus as plans for interplanetary human space missions solidify. […]”
6/ We ran several experiments to test the pre-registered predictions of our account and also compared them to 10 alternative models/processes.
Across these studies, we found consistent evidence that people engage in a process of value-guided construal.
3/ Since Newell & Simon’s time, a lot of work has examined how we simulate actions to plan (part 2), but there’s still a lot unknown about how we represent problems to plan (part 1)
Hence the question “How do we represent problems when planning?”
New preprint with R Dubey, H Mehta and T Griffiths! We analyze Aha moments 🤔💡 as *meta-cognitive prediction errors* to account for key findings in the literature. We also ran a big anagram study! Can you unscramble MARNOD🤔? No? Read the paper and maybe you'll figure it out😜
A computational theory of Aha! moments! New preprint from our lab on why Aha! moments occur and why they feel so good. TLDR: They are a form of meta-cognitive prediction errors i.e. they occur when we surprise ourselves about our own abilities!
7/ In the supplement, we discuss algorithms for construal optimization, and also report several *construal modification* variants of the model designed to capture, at variable levels of resolution, the fine-grained dynamics of the construal process (see image).
Super fun and stimulating talk by
@Mark_Ho_
() today in our lab meeting! Mark builds computational models of social reasoning. He talked about how doing and showing differ, and how Saul Steinberg paints a better picture of planning than google maps does.
What happens when you put your foot 🦶 in your mouth 😮? 🤔
Check out our
#cogsci2019
preprint on unintentional speech acts! We model the role of false assumptions about common ground in
#awkward
situations (eg faux pas).
w J Korman & T Griffiths
Why do initially rewarding things eventually lose their luster? Why does the hedonic treadmill exist?🏃
This fascinating project led by Rachit Dubey uses RL and reward design to understand such vexing existential questions!
Habituation and comparisons can result in depression, materialism, and overconsumption. Why are these disruptive features even part of human cognition?
New paper with Rachit Dubey and Peter Dayan in
@PLOSCompBiol
provides a RL perspective on this question
Really cool paper by
@ManasiMalik
and
@leylaisi
on how human social inferences can be captured by NNs that represent relational visual info!
Interesting to think about as a computationally cheaper alternative to explicit theory of mind
Is purely visual information sufficient to explain human social interaction judgments?
See new preprint w
@leylaisi
"Relational Visual Information explains Human Social Inference:A Graph Neural Network model for Social Interaction Recognition" to find out!
Starts today at 1pm in MacMillan Hall rm 117!
Some planned topics of discussion:
- rethinking the agent/env boundary 🐭🌍
- filling in the perception-action loop 👀🔄🤘
- interdisciplinary similarities/diffs in conceptualizing agency 🙈🙉🙊
and more!
#RLDM2022
So excited for
#RLDM2022
and Saturday's
#RLasAgency
workshop, organized with the amazing
@aharutyu
+
@dabelcs
🤩🤩
We'll be discussing the limits+possibilities of RL as a model of agency in cog sci/neuro/philosphy/AI with a brilliant group of scholars...
10/ I’m also grateful to
@weijima01
,
@RedmondOConnell
and the anonymous reviewer for their incisive reviews. The feedback we received during the review process at
@Nature
was crucial for sharpening our thinking about this project and greatly expanding its scope.
This is a really exciting line of work. It's not only an impressive experimental accomplishment (90 games!), it also does a deep dive into the consequences that rich conceptual models (theories) of tasks can have on how humans learn, plan, and explore.
I’m incredibly grateful to the
@CogCompNeuro
#ccneuro22
community and organizers for giving us the opportunity to share our work and creating such a dynamic and interdisciplinary space to do science
Can’t wait for
#ccneuro23
😄
Also, a special shout-out to the GAC3 team …
Excited to be in SF for jam-packed
@CogCompNeuro
2022!
First, I’ll be presenting recent work w Jon Cohen and
@cocosci_lab
on *rigidity in planning* 🤔💭
Poster P1.24: Tonight at 7:30
Talk: Tomorrow at the 10:30-12 session
Really exciting to be a part of this project, spearheaded by
@dabelcs
!
Dave's thread is a great general overview of the paper. Here, I'll share some thoughts on why this work is relevant for cognitive scientists/psychologists/neuroscientists (short 🧵)
Really excited to share our upcoming
#NeurIPS2021
paper! "On the Expressivity of Markov Reward"
Joint with
@wwdabney
,
@aharutyu
,
@Mark_Ho_
,
@mlittmancs
, Doina Precup, and Satinder Singh.
We ask: What can Markov reward express?
(1/5; a quick 🧵--> )
I’ve always wondered whether resource rationality could be the key to unlocking a number of puzzles in moral psychology/philosophy. Really excited to dive into this!!
“Resource-rational contractualism: A triple theory of moral cognition.” With my incredible co-authors
@NickJChater
, Josh Tenenbaum, and
@fierycushman
. Available here: (1/n)
Our poster is today from 11a-12:40p! We'll be in the official
#CogSci2020
, in the upper-left. To find us: click on one of our names in the "Participants" list, then click "locate".
Poster:
@Mark_Ho_
,
@callfredaway
, Tom Griffiths and I will present work at
#CogSci2020
that models how—and *why*—people decompose big scary tasks 🐲 into smaller, more friendly tasks that make it easier to plan 🧩🐭🐭🐭.
Poster P-3-566, Zoom on Aug 1. Preprint:
To answer this question, we propose that functional fixedness occurs bc people try to minimize cognitive effort by doing two things:
1. Simplifying their representation of a problem, and
2. Sticking with the same kind of simplification on new problems
Tired of social distancing and lockdown? Then check out our
#cogsci2020
workshop on Wednesday! Can't attend live because of timezone/childcare? 90% of talks are already available for asynchronous viewing on our website
#cogcultureworkshop
I largely agree with these points
It takes less mental effort to do 100% of a task yourself than to continuously monitor a machine that does 99% of a task perfectly but fails catastrophically on the last 1% (esp if the 1% is unpredictable!)
It’s unrealistic to think that AI systems will always work perfectly out of the box! 🤖❌😵
This project led by the brilliant
@TheAndiPenguin
proposes a user-centric approach for diagnosing failures when things go wrong, so they can be fixed! 🤖✅🤗
I'm excited to share our
#ICML2023
paper: we develop a user-informed framework for eliciting feedback to diagnose and fix policy failures.
Project page: . [1/8]
Our paper is out in Cognitive Science! We (Robert Walter-Terrill,
@JaraEttinger
, Brian Scholl, and I) ask to what extent perception is (or, from an evolutionary perspective, ought to be) truthful.
How do we compare the scale of language learning input for large language models vs. humans? I've been trying to come to grips with recent progress in AI. Let me explain these two illustrations I made to help. 🧵
Although functional fixedness was introduced by Karl Duncker nearly 80 years ago, it has been notoriously difficult to formalize!
As such, I see this work as an exciting step towards understanding this puzzling and pervasive cognitive bias in modern computational terms 🧠💭
Excited for tomorrow's workshop on games in cog sci! My talk (10:30AM EDT/16:30PM CEST) will focus on work with
@cocosci_lab
on games and resource rational planning representations (e.g., task construal ). Looking forward to seeing everyone (virtually)🎮😊
Announcing the first Games for Intelligence workshop () 🎮🤯 at CogSci 2021! This workshop will examine how games can be used as an alternative to traditional experiments, unlocking larger scale data collection and more ecologically grounded behavior.
1/
When I had just started my job at
@NYU_CNS
@NYUPsych
in 2013, I pitched a project to incoming graduate students.
It was a wild new direction for me: using board games to understand how people think ahead.
7.5 years later, we have a preprint.
1/
Excited to be participating in this summer school organized by
@TheCharleyWu
and
@WataruToyokawa
! If you're interested in computational models of social and/or collective behavior, you should apply!! 😁
🚨Call for applications!🚨 The Computational Summer school on Modeling Social and collective behavior (COSMOS), taking place in Konstanz, DE, between 4 - 7 July organised by me and
@TheCharleyWu
. Application deadline: 14 April. Visit👉 🧵1/3
Hello
#ICML2022
Baltimore!!
I’ll be presenting a poster on our construal work at Friday’s Beyond Bayes workshop (P04) 📈
Come find me if you want to chat 😄
1/ I'm excited to share our new paper, now in
@Nature
!
Paper:
PDF:
Summary 🧵:
We present a new theory of problem simplification to answer an old question in cognitive science and AI:
How do we represent problems when planning?
Excited to be in New Orleans for
#NeurIPS22
!
If you want to chat hit me up 🤙🏼 it’s already been such a blast catching up with folks I haven’t seen in person since pre-pandemic (or ever!)
I’m also incredibly grateful for the time I've spent in the Computer Science department at
@FollowStevens
It's been such a supportive place to build a computational cog sci group, and I’m looking forward to continuing the collaborations we started there!
People often respond to bad behaviour by punishing the perpetrator. But what makes punishments effective? Find out in our new paper, forthcoming in Cognition!! 🥳🎉 Work with
@Mark_Ho_
@justin_w_martin
@fierycushman
URL:
Highlights👇🏽 1/n
‘Functional fixedness’ is the tendency to *fixate* on a specific way of thinking about problems (the proverbial hammer), even if it prevents you from finding better solutions, like in this video
(Warning: it looks upsetting at first, but ends happily!!)
Then, we turn to *inverse* models (e.g., human theory of mind), focusing on what cognitive scientists have to say about the basic building blocks of mentalizing as well as how it shapes human communication and teaching. 3/5
If you are thinking of pre-ordering Code To Joy (), now might be a good time. :-) Apparently, it's easier for the publisher to argue for an audiobook version if there's strong pre-order paperback sales. (Not sure the logic there is sound, but there it is.)
The thing about large language models like GPT-3 and Lambda describing the experience of being self-aware is they can also describe the experience of being a squirrel.
Is there any data/analysis from conferences that switched to double-blind peer review? (Eg thinking of
@cogsci_soc
, which switched over a few years ago)
Massive status bias in peer review.
534 reviewers randomized to review the same paper revealing the low status, high status, or neither author. 65% reject low status, 23% reject high status.
Amazing work by Juergen Huber and colleagues.
#prc9
Excited to be part of work on inferring task representations that are 1. temporal and 2. boolean (e.g. as an alternative to Markov rewards) from expert demonstrations. This representation helps with interpretability and learning tasks compositionally.
Spotted a duck-rabbit in the wild
🐰?🦆
“I shall call the following figure, derived from Jastrow, the duck-rabbit. It can be seen as a rabbit’s head or as a duck’s.”
- Ludwig Wittgenstein, Philosophical Investigations, Part II, §xi
#philosophyjokes
Is GPT4 bad at creative tasks (eg indexed by AP English) bc of fundamental limitations on LLMs or bc it’s a product that’s been fine-tuned to not express strong opinions?
want to get into program synthesis but don't know how to started? I wrote a minimalist intro to modern program synthesis that can help you -- from problem formulation to generating code by fine-tuning llm on huggingface.