Sign in

"Kevin" Tingke Shen (沈庭柯)

@shenkev.bsky.social
59 followers 107 following 60 posts

Cognitive neuroscientist by vocation. Scientist scientist by compulsion. Trying to answer big questions one small idea at a time. PhD candidate with Peter Dayan at Max Planck Institute for Biological Cybernetics. Freestyle hiphop dancer, snowboarder.

PostsRepliesMedia
Reposted by "Kevin" Tingke Shen (沈庭柯)
Niklas Buergi @niklasbuergi.bsky.social · 14/10/2026
Very honored that our work was selected for the Paper of the Year Award from the Society for Neuroeconomics! Many thanks to @socforneuroecon.bsky.social for this special recognition, and to co-recipient @drgokhanaydogan.bsky.social and co-authors Arkady & Christian! bsky.app/profile/nikl...
Me excitedly and happily receiving the award :)
141
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 28/09/2026
From time to time I hear colleagues complain "we're not discovering anything new in cognitive science, we're simply using math to redescribe what we already know intuitively". But I think this is the point. Cog Sci like mathematics, is about building a language to describe "the world".
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/09/2026
I feel like there's a planning horizon mismatch - when you work on projects you're thinking about the next n future projects' joint contribution. But reviewers often (?) have a planning horizon of 1, judging your paper's contribution in isolation...
000
Reposted by "Kevin" Tingke Shen (沈庭柯)
Niklas Buergi @niklasbuergi.bsky.social · 09/09/2026
Autism has long been linked to difficulties with Theory of Mind. But what exactly is driving these differences? And what happens in the brain? In our new preprint, we show that adaptive mentalization might be key. A 🧵 #CogSci #CogNeuro osf.io/preprints/ps...
Repeated Rock-Paper-Scissors task used to study adaptive mentalization, alongside a schematic showing increasingly sophisticated reasoning strategies, from responding to past actions to trying to guess the opponent’s prediction about one’s own next action
13715
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/08/2026
I feel like some amount of scientific pessimism in cogneuro is healthy otherwise we'll go on indefinitely "stamp collecting" and uncovering the endless possible q and Ls
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/08/2026
In conclusion, we have a "universal framework" in cogneuro like we do in physics (RL vs LM) - the mathematical machinery is the same, but simply the systems we deal with in cogneuro are much more complex, the state/transitions cannot be specified in every situation.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/08/2026
The problem with RL is the environment is too complex, there is no compact description of L, transition dynamics must be specified locally p(s'|s, a). Furthermore, there's an extremely large set of possible states s and Lagrangian L.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/08/2026
In LM there's a small set of q and L which describe all *types of physics systems of interest. Gravity, electromagnetism, etc.
110
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/08/2026
In Lagrangian mechanics you can specify the transition function globally using a local function L=1/2M(q)qdot^2-V(q) since the dynamics have invariance. In RL you typically don't enjoy this invariance.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/08/2026
In both cases you specify specific states representing your environment s/q. And state transition functions p/L. The RL/Lagrangian machinery spits out optimal/stationery trajectories.
101
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/08/2026
Recently heard some ppl complain that "cognitive/neuroscience doesn't have, and needs, a universal law like physics!" Isn't the "universal" framework basically reinforcement learning? RL is the analog of the Langragian framework in physics.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 10/08/2026
@tomssilver.bsky.social then perhaps you would also be interested in our paper on analogy as amortized model construal? :) arxiv.org/abs/2507.16511
arxiv.org
Analogy making as amortised model construction
Humans flexibly construct internal models to navigate novel situations. To be useful, these internal models must be sufficiently faithful to the environment that resource-limited planning leads to ade...
110
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 10/08/2026
Also non-scientific topics approached scientifically! One of the most memorable evenings was getting dating advice over dinner from a table of senior researchers 😂
010
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 10/08/2026
I always wondered what the point of going to conferences was. Finally figured it out this year at @cogscisociety.bsky.social and CCN: to maximize the number of interesting, fulfilling scientific conversations! Had such a blast in Rio and NYC, will cherish all the convos I had :)
171
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 21/05/2026
I've been reading James Baldwin's work and can't help but make the analogy that he's "the Einstein of art". His framework for the human experience explains so much variance across human societies. Is so concise. So natural and simple, that it feels like a "General Relativity of art".
020
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 19/05/2026
Just as meta-cognition and meta-control are central to human cognition, they may also be key to building more intelligent AI! Our paper provides a guiding framework. Meta-control is currently a focus of our lab - so more meta-control to come in the near future!
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 19/05/2026
There is growing interest in computational psychiatry to frame maladaptive behavior and cognition as failures of self-monitoring and adjustment - we hope to provide theoretical clarity for future works!
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 19/05/2026
We make several moves: 1) propose the meta-MDP arises from "meta-construal" 2) formalize meta-cognition as meta-level perception 3) link neural implementation and algorithm using the notion of genetic vs specific meta-control and 4) argue for model-based meta-control.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 19/05/2026
Meta-control is studied in meta-cognition, cognitive control, meta-reasoning. Recently, as expected value of control and the value of computation. We consider common core problems such as meta-level state representation, generalization, and prospective vs retrospective calculations of meta-policy.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 19/05/2026
osf.io/preprints/ps... My PI Peter Dayan and I just released our preprint proposing a new framework "recursive-MDP" integrating meta-control and meta-cognition! Couched in RL, recursive-MDP presents a new ontology for meta-control - monitoring and adjusting ongoing cognitive processes.
osf.io
OSF
150
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 25/04/2026
As your castle takes form, other kids on the beach might come help you build the castle. Second, if the amorphous pile gets too big, it becomes unwieldy to shape into a castle.
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 25/04/2026
Then you create the sharp edges and flat surfaces with your shovel. The pile turns into object. A castle. This is the process of writing a scientific paper - to solidify your intuitions. There are two reasons to do this.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 25/04/2026
Explaining the scientific process to a non-scientist friend: it's like building sand castles/sculptures. First you plop a pile of a sand down, roughly resembling some shape. This is the brainstorming and discussions, early stages of a project.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 11/04/2026
And in turn: the dichotomy between "question" and "answer" is a rough approximation of the real process. Science is more like crafting a sculpture.
010
Reposted by "Kevin" Tingke Shen (沈庭柯)
Roxana Zeraati @roxana-zeraati.bsky.social · 01/04/2026
Many thanks to @klaus-tschira-stiftung.de and @gso-forresearchers.bsky.social for this wonderful opportunity to study the mechanisms underlying naturalistic decision processes in the context of foraging behavior! We’ll also soon have openings for PhD students, stay tuned!
1010913
Reposted by "Kevin" Tingke Shen (沈庭柯)
Niklas Buergi @niklasbuergi.bsky.social · 09/03/2026
How does the brain decide which mental strategy to use when inferring others' beliefs? Excited to (finally!) see my first first-author paper out @natneuro.nature.com Summary below 🧵 #CogSci #CogNeuro www.nature.com/articles/s41...
nature.com
A neural signature of adaptive mentalization | Nature Neuroscience
18830
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 05/03/2026
After my lab visit in Lausanne, I spent some time in Aosta Valley and Evolène. The scenery was absolutely breathtaking.
020
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
8/ Arguably for all of human history, we've built systems that work given a fixed, narrow construal of a problem (bridges, coffee machines, vacumm robots, CNN image classifier). I believe understanding how agents can construe problems themselves is the key to building general intelligence.
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
7/ In the future, we plan to investigate aspects of this framework such as: 1) how high-level objects or modules emerge from low-level states and policies, 2) what analogy-making processes are employed by humans to retrieve modules from the library, 3) what processes refine the library.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
6/ An agent's library of models/moduels is refined by creating more and more abstract analogies over time. E.g. the abstract "key"-"door" notion, rather than physical key-doors. Then "chocolate is the key to my heart (door)". These abstractions aid in future analogy-making.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
5/ From a meta-cognition and meta-control point of view, analogy-making processes must retrieve amortized models by respecting two types of costs: 1) cost of construal, and 2) cost of finding a solution (planning) given a construal.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
4/ For example, after experience with a physical key-door, an agent reasons by analogy to type a password into a login screen. The optimal policy ("insert password into login") is imported and applied without needing experience with the login screen.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
3/ A good construal reduces the computations required to find/plan a solution/policy (sometimes revealing a trivial policy). But searching for a good construal can be more costly than solving the original problem. Agents should amortize past construals into a library - retrieval is analogy-making.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
2/ Much of RL ignores the "construal problem" whereby state + action space, and allowed state-action mapping is constructed. Also true for many computational frameworks (e.g. hypothesis space in Bayesian inference). But arguably this construal process does "the heavy lifting" for problem solving.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
1/ We bridge the ideas of analogy, amortization and model construal in reinforcement learning. Agents amortize models or submodels ("modules") from previous environments. These modules are retrieved via analogy in new environments, allowing zero-shot transfer for planning or action-selection.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
Possibly the line of work I'm most excited about is my collaboration with @davidnagy.bsky.social, Hanqi Zhou + PIs on the connection between analogies and model-reuse in reinforcement learning. In our recent perspective paper, we chart our future research agenda. arxiv.org/pdf/2507.16511
arxiv.org
110
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
3/ In the bigger picture, complexity is a meta-cognitive measure of representation relative to cognitive limits. Number of segments can't be the full picture. The brain represents other aspects about an image - structural relationship between objects, object affordances, etc.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
2/ Our work is one example of a novel approach to building computational, cognitive models by leveraging outputs of neural networks. Segment Anything Model (SAM) serves as the "low-level perception", and complexity can be computed "on-top" of the outputs of this low-level perceptual system.
110
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
1/ Previous models of image complexity relied on statistical features about edges, colors, etc. We show that a single "object-centric" feature of number of segments detected by SAM predicts image complexity extremely well. This suggests complexity depends on what humans "care about" in the image.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
Understanding "complexity" has been a goal of mine. In @surabhisnath.bsky.social and my paper, we build a novel model of complexity judgments for images using the Segment Anything Model, finding surprisingly high correlation between the number of segments and complexity. arxiv.org/abs/2403.03134
arxiv.org
Simplicity in Complexity : Explaining Visual Complexity using Deep Segmentation Models
The complexity of visual stimuli plays an important role in many cognitive phenomena, including attention, engagement, memorability, time perception and aesthetic evaluation. Despite its importance, c...
141
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
References 1. Wolfram, Stephen. A New Kind of Science. Wolfram Media, 2002. Chapter 12. 2. Newell, Allen. “You Can’t Play 20 Questions with Nature and Win: Projective Comments on the Papers of This Symposium.” Machine Intelligence, Routledge, 2012, pp. 121–146.
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
Awesome visit at @gerstnerlab.bsky.social in EPFL, Lausanne. What a stunning campus too!
140
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
3/ #novelty #exploration #risk-sensitivity #surprise #striatum
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
3/ Dataset paper. pubmed.ncbi.nlm.nih.gov/36130595/
pubmed.ncbi.nlm.nih.gov
Striatal dopamine explains novelty-induced behavioral dynamics and individual variability in threat prediction - PubMed
Animals both explore and avoid novel objects in the environment, but the neural mechanisms that underlie these behaviors and their dynamics remain uncharacterized. Here, we used multi-point tracking (DeepLabCut) and behavioral segmentation (MoSeq) to characterize the behavior of mice freely interact …
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
2/ We capture different behavioral modes of mice exploring a novel object from (Akiti et al. 2022), providing a generative, normative explanation of rational intermittent approach strategies. There is surprising individual differences from "brave" animals to "timid" animals with limited approach!
000
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
1/ The project continues a line of research in Peter's lab modeling risk aversion normatively using Conditional Value at Risk (CVaR), with applications to depression, anxiety. CVaR modifies the objective of the agent, explaining risk aversion in terms of planning focusing on negative outcomes.
010
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 03/03/2026
Last year my PI Peter Dayan and I published our paper on risk-sensitive exploration in eLife! TD;LR AI agents can be purely neophilic since they can "reset". Ample evidence of how biological agents trade off neophobia vs neophilia. We built a computational model! elifesciences.org/reviewed-pre...
elifesciences.org
Risking your Tail: Modeling Individual Differences in Risk-sensitive Exploration using Bayes Adaptive Markov Decision Processes
451
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 02/03/2026
16/16 Maybe the problem is us Maybe describing cognition is out of reach for scientists who are cognitively limited themselves. We are “not the lathe that can make itself”. Yet we tread on...
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 02/03/2026
15/16 Unified model of cognition …is not on the horizon. In both LLM and RL/PI cases, we never beat Newell’s challenge. We made progress by 1) engineering systems that work but are equally hard to describe with theory, 2) merging results into frameworks - complex behaviors also evade description.
100
"Kevin" Tingke Shen (沈庭柯) @shenkev.bsky.social · 02/03/2026
14/16 Generative models RL and PI differ from bags of heuristics in that they allow scientists to build generative models of behavior where mechanisms interact. However the behaviors generated by these models can be too complex for the cognitive scientist to explain succinctly (find a shortcut).
100