Sign in

Will Redman

@wtredman.bsky.social
157 followers 264 following 74 posts

Assistant professor JHU ECE | Dynamics of learning and learning of dynamics | He/Him dynamical-intelligence-group.github…

PostsRepliesMedia
Will Redman @wtredman.bsky.social · 21/09/2026
∂DSA finds lots of RNN solutions to the same integration task: Line attractors ✅ Feedforward chains ✅ Oscillations ✅ Unstable dynamics ✅
021
Will Redman @wtredman.bsky.social · 21/09/2026
DSAmaxxing for comparing dynamical systems! Some really great work from @neurostrow.bsky.social on advancing DSA to be able to shed light on complex, noisy neural dynamics AND using DSA as a regularizer to discover novel RNN solutions Really lucky to have been included and involved in this!
082
Reposted by Will Redman
Caswell Barry @caswell.bsky.social · 19/09/2026
Masa's paper is out in Hippocampus. The left-right alternation of hippocampal theta sequences at T-maze choice points, much loved as a 'planning' signal, falls straight out of MEC dynamics. Well done @masahironakano.bsky.social, fun project with @clopathlab.bsky.social doi.org/10.1002/hipo.70131
04615
Will Redman @wtredman.bsky.social · 08/09/2026
Very cool!
011
Reposted by Will Redman
William Gilpin @wgilpin.bsky.social · 08/09/2026
Our group discovered that reasoning models produce fractals when asked to solve hard problems. We can use nonlinear dynamics to probe the thinking processes of recurrent depth models on Sudoku, mathematics, and even ARC-AGI (1/N) arxiv.org/abs/2609.04963
24010
Reposted by Will Redman
Ken Miller @kenmiller.bsky.social · 26/08/2026
This is wild. >1/2 of HC synapses disappear in artificial hibernation, come back in same/similar locations & memories intact. ~1/2 of engram synapses-between two memory-tagged cells-are clustered on dendrites, those preferentially spared. Not biggest synapses. 1/ www.science.org/doi/10.1126/...
This shows the visual abstract and a paragraph from the extended abstract of the linked article.
110534
Reposted by Will Redman
Ben Hayden @benhayden.bsky.social · 13/08/2026
New paper from the BCM Neurosurgery research team! "Neural basis of compositional control" led by @assiachericoni.bsky.social and @justfineneuro.bsky.social ! 🧵 www.nature.com/articles/s41...
nature.com
Neural basis of compositional control - Nature
Behaviour of human participants in a prey-pursuit task reflects dynamic blending of goal-specific control policies, with hippocampus estimating latent states, anterior cingulate cortex orchestrating p...
39546
Will Redman @wtredman.bsky.social · 29/07/2026
Congrats @xiaoxiao-lin.bsky.social !
010
Reposted by Will Redman
Satpreet (Sat) Singh @satpreetsingh.bsky.social · 29/07/2026
How do you "see" with electric eyes? How does collective behavior emerge from individual interactions? Nocturnal weakly electric fish evolved to do this, but studying naturalistic social behavior is very hard. Our solution? Virtual 'fish' 🤖🐟⚡ 📄 arxiv.org/abs/2511.08436
17925
Reposted by Will Redman
Hugo Spiers @hugospiers.bsky.social · 16/07/2026
Is space & time coding in entorhinal cortex interchangeable? Cells switching between space & time as needed by task? For some cells yes, and other no: Spatial and temporal representations are organized along a stable coding gradient in the medial entorhinal cortex www.biorxiv.org/content/10.6...
biorxiv.org
13511
Reposted by Will Redman
Michael Goard @michaelgoard.bsky.social · 13/07/2026
New review on ecological visual processing. Really enjoyed writing this as part of a longstanding collaboration with @crisniell.bsky.social, @mbeyeler.bsky.social, and @spencerlaveresmith.bsky.social
092
Reposted by Will Redman
Fatih Dinc @fatihdinc.bsky.social · 23/06/2026
Why did RNNs fail to learn long-term dependencies? What if we added one more modification, maybe now? It turns out we can give a pretty broad, analytical answer! See the attached paper for a rigorous treatment using centre manifolds, low-rank RNNs, and dynamical systems theory! go.aps.org/4fXWEeF
go.aps.org
Ghost Mechanism: An Analytical Model of Abrupt Learning in Recurrent Networks
This study establishes the ghost mechanism as an underlying mechanism for abrupt learning, whereby the recurrent neural network develops ghost points---transient dynamical bottlenecks---and identifies...
03510
Will Redman @wtredman.bsky.social · 22/06/2026
Woah!
010
Reposted by Will Redman
andy alexander @andyalexander.bsky.social · 19/06/2026
@fatihdinc.bsky.social won the Richard C. DiPrima Prize from SIAM! He is an incredible scientist and I feel genuinely lucky to work with him alongside @wtredman.bsky.social and @ninamiolane.bsky.social . Congrats Fatih! Read more here: news.ucsb.edu/2026/022632/...
news.ucsb.edu
New mathematical frameworks reveal how stable thoughts emerge from chaotic brain activity
Postdoc Fatih Dinc has been recognized by the the Society for Industrial and Applied Mathematics for his work developing a bridge between mathematics and neuroscience.
092
Will Redman @wtredman.bsky.social · 16/06/2026
Sheesh!
010
Will Redman @wtredman.bsky.social · 05/06/2026
Really cool work!
120
Reposted by Will Redman
Roddy Grieves @roddy-grieves.bsky.social · 05/06/2026
Ever wonder how your brain navigates the real world? 🧠 Most spatial navigation studies use flat, 2D mazes. But the real world is hilly, irregular, and bumpy! ⛰️ Our new paper in #ScienceAdvances asked a simple question: How does the brain map uneven terrain? 📄 doi.org/10.1126/scia... 🧵👇️️️ 1/
26018
Reposted by Will Redman
Tamir Eliav @tamir-eliav.bsky.social · 29/05/2026
🚨 New paper in @nature.com We asked why two hippocampal areas with very different anatomy, CA3 and CA1, often seem to code space so similarly? By recording bats flying up to 200m, we found that the difference was hidden by scale! www.nature.com/articles/s41... 🧵 1/11
nature.com
Sparse-to-dense coding transformation between hippocampal areas CA3 and CA1 - Nature
The hippocampus exhibits a CA3-to-CA1 coding transformation that combines fast learning with an efficient, compressed neural code.
19533
Reposted by Will Redman
David G. Clark @david-g-clark.bsky.social · 06/05/2026
Now in PRE: "Transient dynamics of associative memory models." I argue that the "blackout catastrophe" is not catastrophic when viewed from an out-of-equilibrium, dynamical perspective. Journal: journals.aps.org/pre/abstract/10.1103/42y2-bsh1 PDF: dclark.io/media/clark-...
journals.aps.org
Transient dynamics of associative memory models
Associative memory models such as the Hopfield network and its dense generalizations with higher-order interactions exhibit a ``blackout catastrophe''---a discontinuous transition where stable memory ...
1198
Reposted by Will Redman
andy alexander @andyalexander.bsky.social · 28/04/2026
Very excited to announce the first preprint from the lab @ucsantabarbara.bsky.social -- "Predictive pursuit emerges in high-dimensional RNNs." @wtredman.bsky.social (starting his own group at JHU this fall) summarizes the work nicely in this thread.
13510
Will Redman @wtredman.bsky.social · 28/04/2026
And I’m very happy to discuss more so feel free to reach out (wredman4@jh.edu) with any comments/questions/critiques! 24/
010
Will Redman @wtredman.bsky.social · 28/04/2026
There’s a lot more to do and a lot of ways to extend/expand the computational results. One of which is to see if/how the use of RL (like @satpreetsingh.bsky.social 's work nature.com/articles/s42...), instead of supervised learning, changes things 23/
nature.com
Emergent behaviour and neural dynamics in artificial agents tracking odour plumes - Nature Machine Intelligence
Olfactory navigation is a well-studied topic in insect behaviour, but many aspects of the challenging task of odour plume tracking are unknown. In a deep reinforcement learning approach, artificial ag...
120
Will Redman @wtredman.bsky.social · 28/04/2026
And emphasize that, while egocentric target cells may play an important causal role in pursuit, allocentric information about the target (e.g., “social place cells” in hippocampus) may be critical for enabling predictive pursuit 22/
110
Will Redman @wtredman.bsky.social · 28/04/2026
Our work provides an ethologically motivated example of where high-dimensional representations enable different emergent behavior than low-dimensional representations 21/
130
Will Redman @wtredman.bsky.social · 28/04/2026
For all RNNs (low- and high-rank), we found that egocentric target information could be decoded with similar accuracy. However, allocentric info was better decoded in high-rank RNNs and better decoded when considering the prospective ("future") position of the target 20/
110
Will Redman @wtredman.bsky.social · 28/04/2026
What does the extra capacity enabled by increased rank buy the RNNs? We performed linear decoding of different features of the task from the population activity and compared decoding accuracy across ranks 19/
110
Will Redman @wtredman.bsky.social · 28/04/2026
However, when we did this, we found that the predictive behavior of the RNNs was greatly decreased (despite similar pursuit performance)! Increasing the rank of the RNN led to an increase in predictive behavior and an increase in population activity dimensionality 18/
110
Will Redman @wtredman.bsky.social · 28/04/2026
A natural approach for addressing this is to train low-rank RNNs (i.e., RNNs with the same number of units, but constrained to have only a few degrees of freedom in the way they are connected) 17/
110
Will Redman @wtredman.bsky.social · 28/04/2026
Having shown that the RNN learns to pursue predictively, and found evidence that it is able to generate and maintain an internal model of the target’s dynamics, we asked what computations support this? 16/
110
Will Redman @wtredman.bsky.social · 28/04/2026
This behavior was striking, so @xiaoxiao-lin.bsky.social trained mice to pursue a moving laser pointer in an environment with similar “periodic boundaries”. We find the mice learn a similar strategy to the RNNs (see figure above) 15/
110
Will Redman @wtredman.bsky.social · 28/04/2026
We also train the RNN in an environment with “periodic boundaries” and find that, despite the non-Euclidean geometry, the RNN learns to make a direct path to the wall where the target will appear and then waits 14/
110
Will Redman @wtredman.bsky.social · 28/04/2026
We find that, even when info about the target is missing 50% of the time, the RNN can still achieve good pursuit performance and generate trajectories that are significantly more predictive than the non-predictive control model 13/
110
Will Redman @wtredman.bsky.social · 28/04/2026
To further explore the ability of the RNNs to generate and update internal models that can be leveraged for prediction, we train RNNs to pursue the target, with some of the inputs about the target “masked out” 12/
110
Will Redman @wtredman.bsky.social · 28/04/2026
Across several different metrics, we find that the RNN’s trajectories are significantly different from the non-predictive control model's and are significantly more aligned to an “optimally predictive” trajectory 11/
110
Will Redman @wtredman.bsky.social · 28/04/2026
But does the RNNs learn to pursue the target predictively? Or are they just reactively following the target? To test this, we compare the trajectories of the RNN to trajectories of a non-predictive (“reactive”) control model 10/
110
Will Redman @wtredman.bsky.social · 28/04/2026
We find units with similar tuning in the RNN. Removing them leads to a significant decrease in pursuit accuracy, relative to random ablations. In addition, removing the units with the least egocentric target tuning has a significantly smaller impact 9/
110
Will Redman @wtredman.bsky.social · 28/04/2026
Electrophysiological recordings have found neurons in RSC and PPC that are tuned to the egocentric location of the target (“egocentric target cells”) cell.com/cell-reports... cell.com/cell-reports... biorxiv.org/content/10.6... 8/
cell.com
Adaptive integration of self-motion and goals in posterior parietal cortex
Alexander et al. examine rats pursuing visual targets and characterize emergent predictive behaviors. Relative to free exploration, pursuit elicits enhanced coding of self-motion in the parietal corte...
110
Will Redman @wtredman.bsky.social · 28/04/2026
Inspired by Andy’s experiments, we include trials where the target moves along pseudo-randomly trajectories (RTs) and along characteristic trajectories (CTs). We find the RNN develops complex pursuit behavior, performing well on RTs and CTs 7/
110
Will Redman @wtredman.bsky.social · 28/04/2026
To provide a complementary approach to experiments, we developed and trained RNN models (using supervised learning) to pursue a moving target 6/
110
Will Redman @wtredman.bsky.social · 28/04/2026
While there has been foundational work in this area (e.g., nature.com/articles/s41...), there is still a lot that is not well understood about the neural computations that support predictive pursuit 5/
nature.com
The neural basis of predictive pursuit - Nature Neuroscience
Yoo and colleagues find that while pursuing virtual prey, monkeys predict the prey’s upcoming movements, and neurons in the dorsal anterior cingulate cortex tracked prey position, velocity and acceler...
130
Will Redman @wtredman.bsky.social · 28/04/2026
Andy developed a powerful experimental paradigm to study pursuit in the lab and found that rats could learn to exploit repeated structure of the target’s dynamics to make anticipatory trajectories cell.com/cell-reports... 4/
110
Will Redman @wtredman.bsky.social · 28/04/2026
Tracking moving objects is ethologically important for many organisms and a wide range of species can do it predictively (e.g., bats, rodents, non-human primates) 3/
110
Will Redman @wtredman.bsky.social · 28/04/2026
TLDR (see 🧵 for motivation and details): We train RNNs to pursue dynamically moving targets and find that predictive behavior emerges only when the model is high-rank, despite low-rank models having similar pursuit performance 2/
130
Will Redman @wtredman.bsky.social · 28/04/2026
New preprint out 🚨 “Predictive pursuit emerges in high-dimensional recurrent neural networks”! This was an awesome collaboration with co-first author @fatihdinc.bsky.social , @andyalexander.bsky.social, @xiaoxiao-lin.bsky.social, and May Chen biorxiv.org/content/10.6... 1/
biorxiv.org
12912
Reposted by Will Redman
Kanaka Rajan @kanakarajanphd.bsky.social · 22/04/2026
✍️ In the @kempnerinstitute.bsky.social blog: our new tool built to compare the dynamics of complex systems when both internal circuitry and external inputs shape their behavior. Catch @annhuang42.bsky.social presenting this work at #ICLR! kempnerinstitute.harvard.edu/research/dee...
kempnerinstitute.harvard.edu
InputDSA: Demixing then comparing recurrent and externally driven dynamics in complex systems - Kempner Institute
We explored how to measure the similarity between two complex systems when they are driven by external inputs, like biological neural circuits or reinforcement learning agents. Our novel method, calle...
14315
Reposted by Will Redman
Ann Huang @annhuang42.bsky.social · 23/04/2026
[ #ICLR2026 ] How do we know if two systems are performing the same computation when they are constantly driven by different external inputs? 🧠🤖 I’ll be presenting our novel method InputDSA tomorrow April 23 (2:15pm-4:45pm EDT in Pavilion 3 P3-#1614)📍 Come swing by our poster! I’d love to chat!
0248
Will Redman @wtredman.bsky.social · 23/03/2026
Excited to be at Max Planck Institute for Intelligent Systems for @cpalconf.bsky.social in Tübingen, Germany! A beautiful setting for a great 3 days ahead!
041
Will Redman @wtredman.bsky.social · 12/03/2026
Super excited about this work and to have gotten a chance to work with @andyalexander.bsky.social @fatihdinc.bsky.social @xiaoxiao-lin.bsky.social and May Chan on predictive pursuit in RNNs and mice!
082
Reposted by Will Redman
Kempner Institute at Harvard University @kempnerinstitute.bsky.social · 09/03/2026
NEW from the #KempnerInstitute: InputDSA, a tool to separate intrinsic dynamics from input-driven effects—enabling accurate, efficient comparisons of complex systems with external inputs. Read the #DeeperLearning blog post by @annhuang42.bsky.social & @kanakarajanphd.bsky.social: bit.ly/4bkrvy9
bit.ly
InputDSA: Demixing then comparing recurrent and externally driven dynamics in complex systems - Kempner Institute
We explored how to measure the similarity between two complex systems when they are driven by external inputs, like biological neural circuits or reinforcement learning agents. Our novel method, calle...
0219
Reposted by Will Redman
Transactions on Machine Learning Research @tmlrorg.bsky.social · 05/03/2026
Wanna earn yourself one of these bad boys? Sign up to be an Action Editor for TMLR! We are also looking for reviewers as well. Please fill out one of the following forms: - Volunteer to action editor: docs.google.com/forms/d/e/1F... - Volunteer to review: docs.google.com/forms/d/e/1F...
082