Similar visual tuning conceals different recovery dynamics in a model of macaque V1
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Similar visual tuning conceals different recovery dynamics in a model of macaque V1".
Marcus: Here's the gist of the paper. Biologically detailed neuronal networks can model how brain circuits perform observed functions, but their high computational cost limits the comparisons needed to test biological hypotheses The smooth deep surrogate of local responses provides derivatives, allowing us to analyze linear stability and to trace how a perturbation in brain activity is transmitted through different pathways Our approach offers efficient evaluation with an interpretable circuit structure, making systematic comparisons across stimuli and biological hypotheses practical Model Construction and Multiscale Framework The multiscale modeling approach treats neuronal activity at different anatomical scales separately, using results from one scale to inform computations at another This involves computing the response of small local neuronal circuits in advance and reusing these responses across a larger cortical circuit Local dynamics are represented by three local response functions, denoted as ΨS, ΨC, and ΨI, which map inputs to mean firing rates for simple (S), complex (C), and inhibitory (I) cells within a pixel.
Ines: These local responses are approximated by parameterized deep neural networks (DNNs) trained on five inputs: orientation coordinate, contrast, L6 drive, recurrent excitation, and recurrent inhibition Coarse-grained dynamics involve interactions between local populations that determine the L4 and L6 input rates based on spatial connectivity kernels.
Marcus: Information passes between scales as population activity determines input, and local neuronal responses determine population activity as the output, evolving the whole circuit for one step on time.
Ines: Computational Acceleration and Differentiability The use of deep surrogates replaces the lookup and interpolation step with a DNN approximation, leading to significant speedups in steady-response calculations These learned functions are smooth, providing derivatives that describe how small changes in a population’s inputs change its firing rate.
Paper summary: Ines: So to recap this paper, "Similar visual tuning conceals different recovery dynamics in a model of macaque V1", they are using this multiscale approach with deep neural networks to speed up simulations and then using those smooth local responses to analyze stability and how signals travel through the circuit.
Marcus: They start by saying that the macaque visual system already shows diversity in how layer four C alpha neurons respond to different stimulus orientations, which is something we can measure at this early cortical stage according to their work on "Similar visual tuning conceals different recovery dynamics in a model of macaque V1" <ref:2611.13692#pg2>.
Yuki: That diversity—the spatial and temporal-frequency tuning—is what makes the circuit models interesting because they need to reproduce all these visual properties simultaneously using just one set of equations and parameters across many stimuli, according to the paper.
Ines: The core claim is that even when different circuit pathways produce similar steady responses to a stimulus, those pathways can have very different recovery dynamics after a perturbation.
Marcus: They find that steady responses and how the circuit recovers from a change are two complementary constraints on what we think a candidate circuit might be.
Yuki: So it's not just about getting the right firing rate map; it’s about the transient behavior, like how long it takes to settle down after a visual flicker or change in light.
Conclusion: Ines: So to wrap up, the authors of "Similar visual tuning conceals different recovery dynamics in a model of macaque V1" are showing that just looking at the steady response isn't enough to tell if two circuits are really similar.
Marcus: They show that you can have different circuit pathways that all produce the same firing rate map for a visual stimulus, but those pathways will recover from a perturbation in very different ways.
Yuki: This is important because it means we need to measure both how the circuit fires and how it recovers to actually distinguish between competing biological hypotheses about what's happening in the brain.
Ines: They’ve done this by using steady responses and perturbation recovery together, which gives us a way to test whether similar visual functions are supported by similar underlying mechanisms.
Marcus: The computational speed-up they achieved is also notable, estimating an order one hundred thousand advantage over spiking simulations after they built their library and trained their deep neural networks.
Yuki: It opens the door for testing many different combinations of pathway strengths against real biological data, which is a big step forward in understanding visual processing.
Xialei Huang, Qingyu Yan, Zheng Bian, Zhuo-Cheng Xiao
Division of Art and Science, NYU Shanghai · Applied Physics Graduate Program, Rice University · Instituto de Matemática, Estatística e Ciência da Computação da Universidade de São Paulo · Department of Mathematics, Imperial College London · NYU-ECNU Institute of Mathematical Sciences at NYU Shanghai · NYU-ECNU Institute of Brain and Cognitive Science at NYU Shanghai
q-bio.NC, q-bio.QM
Submitted: 2026-11-10
Updated: 2026-11-10
Code: https://github.com/XialeiH/Spectral_
License: http://creativecommons.org/licenses/by-nc-sa/4.0/
Importance score: 83/100
The gist: Biologically detailed neuronal networks can model how brain circuits perform observed functions, but their high computational cost limits the comparisons needed to test biological hypotheses The…
Key concepts
- Multiscale Modeling
- This approach treats neural activity at different spatial scales separately. It involves calculating the response of small local circuits first, then using those results to model interactions across a larger cortical area. This helps capture how local details influence global circuit behavior.
- Deep Surrogates
- Instead of running slow, detailed spiking network simulations for every calculation, deep neural networks are trained to quickly approximate the response of small local circuits. These learned functions are smooth and provide derivatives, allowing researchers to analyze stability and how activity changes over time efficiently.
- Complementary Constraints
- The paper shows that steady-state responses and perturbation recovery dynamics provide different, yet necessary, constraints on candidate brain circuits. Pathway combinations might look similar in terms of their stable firing rate but behave differently when disturbed, highlighting the need to study both aspects together.
Summary
Biologically detailed neuronal networks can model how brain circuits perform observed functions, but their high computational cost limits the comparisons needed to test biological hypotheses The smooth deep surrogate of local responses provides derivatives, allowing us to analyze linear stability and to trace how a perturbation in brain activity is transmitted through different pathways Our approach offers efficient evaluation with an interpretable circuit structure, making systematic comparisons across stimuli and biological hypotheses practical
Model Construction and Multiscale Framework
The multiscale modeling approach treats neuronal activity at different anatomical scales separately, using results from one scale to inform computations at another This involves computing the response of small local neuronal circuits in advance and reusing these responses across a larger cortical circuit Local dynamics are represented by three local response functions, denoted as ΨS, ΨC, and ΨI, which map inputs to mean firing rates for simple (S), complex (C), and inhibitory (I) cells within a pixel. These local responses are approximated by parameterized deep neural networks (DNNs) trained on five inputs: orientation coordinate, contrast, L6 drive, recurrent excitation, and recurrent inhibition Coarse-grained dynamics involve interactions between local populations that determine the L4 and L6 input rates based on spatial connectivity kernels. p. 5 parameterized deep neural networks predict the S, C, and I rates. Each receives five inputs: an
orientation coordinate defined relative to the local LGN alignment, contrast, L6 drive, recurrent
excitation, and recurrent inhibition. The local response functions Ψ
then supplies new rates for those inputs. Thus information passes between the two scales: population
activity determines input, and local neuronal responses determine the population activity as the
output, evolving the whole circuit for one step on time.Where the paper says it
Where the paper says it
Computational Acceleration and Differentiability
The use of deep surrogates replaces the lookup and interpolation step with a DNN approximation, leading to significant speedups in steady-response calculations These learned functions are smooth, providing derivatives that describe how small changes in a population’s inputs change its firing rate. This differentiability allows for the analysis of linear stability and to trace how a perturbation in brain activity is transmitted through different pathways The combined acceleration from coarse-graining and deep surrogates yields an estimated evaluation advantage of order 105 over spiking simulation after library construction and training. p. 0 In particular, we approximate these local responses by deep neural networks and achieve
an estimated acceleration of the order 105
for steady-state evaluation relative to spiking network
simulation. Beyond computational acceleration, the smooth deep surrogate of local responses
provides derivatives, allowing us to analyze linear stability and to trace how a perturbation
in brain activity is transmitted through different pathways.Where the paper says it
Stability Analysis via Perturbation Theory
The linearized dynamics near the steady state are described by the Jacobian matrix J = DfΦ(f∗), which describes the recurrent response to a small activity change. The leading right singular vector is identified as the perturbation direction that recruits the largest recurrent response, while its time course includes population relaxation and repeated feedback Eigenvalues with real parts just below one describe slowly decaying or persistent patterns, whereas a real part above one implies instability of f∗. p. 6 The amplitude decays when the real part of λ is below one and
grows when it is above one. Therefore, eigenvalues with real parts just below one describe slowly
decaying or persistent patterns, whereas a real part above one implies instability of f
∗
.Where the paper says it
Pathway Constraints on Dynamics
The analysis examines how distinct circuit pathways shape these dynamics by using fixed-point-preserving (FPP) interventions to compare pathway contributions at a common steady response L6 feedback supports slow recovery, while recurrent inhibition stabilizes the response. The analysis shows that freezing L6 feedback moves the leading eigenvalue away from the stability boundary, and freezing inhibitory input exposes an unstable excitatory subsystem Coordinated changes in pathway strengths can approximately preserve aggregate E/I firing-rate maps and broad orientation tuning while changing transient amplitudes and slow recovery times. p. 7 Freezing L6 feedback moves
the leading eigenvalue away from the stability boundary (β6 = −1, Fig. Freezing inhibitory
input has the opposite effect: originally negative growth rates become positive, corresponding to
eigenvalues of J beyond that boundary (βI = −1, Fig.Where the paper says it
Complementary Constraints from Steady Responses
The paper demonstrates that steady responses and perturbation recovery place complementary constraints on candidate circuits Coordinated pathway changes can preserve broad visual tuning while changing recovery dynamics, which distinguishes pathway combinations that produce the same steady response. Joint measurements of steady responses and perturbation recovery could therefore distinguish pathway combinations that satisfy the same steady-response criterion alone The compensation results show why dynamical distinctions matter for circuit inference. p. 1 In V1, coordinated
changes in recurrent excitation, inhibition, and feedback from other layers may approximately
preserve a steady firing-rate map while substantially changing transient responses. Comparing both
NESS and recovery dynamics could therefore distinguish pathway combinations that NESS alone
leave unresolved.Where the paper says it
Experimental Validation and Future Directions
The model's predictions regarding inhibitory stabilization and L6-supported persistence can be tested using high-density laminar electrophysiology in macaque V1 Pathway contributions could be examined by modulating inhibitory drive within L4 and the activity of L6 using GABAergic optogenetic manipulation Future directions suggest testing whether pathway relationships persist across spatial frequency, temporal frequency, and motion direction as well as orientation and contrast. The same approach can be extended to multiple V1 layers and higher visual areas. p. 1 To answer these questions, we perform perturbation
analysis near NESS. Compared to costly animal experiments, which usually require simultaneous
recordings across multiple areas 15–17, our computational approach can be implemented in an efficient,
biologically realistic circuit model.Where the paper says it
Methodological Details
The model organization includes N = 40 × 40 grid of pixels, each containing 40 cells divided into S, C, and I populations The local response function is determined by the stimulus-dependent LGN drive and three cortical input channels: L6 feedback, L4 recurrent excitation, and L4 recurrent inhibition. The DNN surrogate approximates each library response function ΨQ by a separately parameterized DNN Ψb Q with 19,297 trainable parameters Equilibrium computation solves Φ(f) − f = 0, and the stability criterion requires all eigenvalues of L to have negative real parts. p. 6 The vertical
line at one in the spectral plots is thus the stability boundary of the rate model. The parent CG model and DNN surrogate show similar spectral organization: many eigenvalues
lie near zero, while a smaller group extends towards the stability boundary (Fig.Where the paper says it
Results and Interpretation
The analysis reveals that the spatial structure of the strongest recurrent responses closely resembles that of the inhibitory pathway, whereas the slow patterns resemble those supported by L6 feedback The leading eigencluster envelopes show where the leading patterns are expressed, and their cosine similarity compares support rather than principal angles between invariant subspaces. The readout in Fig. p. 9 Perturbations of the firing-rate map also produce
transient errors in the orientation read out from the population, even while the stimulus remains
fixed (Fig. We measure small orientation shifts around the reference stimulus with a readout
that removes first-order gain and contrast changes (Methods). Consistently, most of the decoding deviation during recovery
comes from the preferred and neighboring domains, while the orthogonal-domain contribution is
small. The spatial profiles and readout thus suggest that the tested disturbances mainly distort the
representation near the presented orientation, rather than strongly recruit an orthogonal-orientation
pattern. Together, pooling and pathway cancellation help explain why the recurrent circuit responds
weakly to many population changes. We note that pathway effects cancel for selected activity
perturbations at fixed circuit parameters, whereas the later compensation analysis varies pathway
strengths together to preserve steady responses approximately.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion and Significance
Perturbation dynamics provide constraints on cortical circuits that steady visual responses alone leave unresolved Joint steady-response and perturbation measurements could therefore distinguish pathway combinations that satisfy the same steady-response criterion The paper concludes that joint measurements of steady responses and perturbation recovery could distinguish pathway combinations that remain similar under the steady-response criterion alone. The computational scope is estimated at order 105 over spiking simulation after library construction and training The common model time scale is held fixed across pathway comparisons, allowing for a direct comparison of recovery dynamics. p. 5 Complete equilibrium calculations are also faster across the tested
field sizes (Fig. Combined with the earlier 103 CG acceleration, this gives an estimated
evaluation advantage of order 105 over spiking simulation after library construction and training
(Methods). Similar visual tuning conceals different recovery dynamics in a model
of macaque V1
Xialei Huang1
, Qingyu Yan2
, Zheng Bian3,4,∗
, and Zhuo-Cheng Xiao1,5,6,∗
1Division of Art and Science, NYU Shanghai, Shanghai 200124, China
2Applied Physics Graduate Program, Rice University, Houston, TX 77005, USA
3
Instituto de Matemática, Estatística e Ciência da Computação da Universidade de São Paulo, São Paulo,
SP 05508-090, Brazil
4Department of Mathematics, Imperial College London, London SW7 2AZ, UK
5NYU-ECNU Institute of Mathematical Sciences at NYU Shanghai, Shanghai 200124, China
6NYU-ECNU Institute of Brain and Cognitive Science at NYU Shanghai, Shanghai 200124, China
∗Corresponding authors
Abstract
Biologically detailed neuronal networks can model how brain circuits perform observed
functions, but their high computational cost limits the comparisons needed to test biological
hypotheses. To address this limitation, we extend a multiscale approach in which responses of
small populations of spiking neurons are computed in advance and reused across a larger cortical
circuit.Where the paper says it
Where the paper says it
Methods Summary
The model organization inherits the local-population construction and spatial interactions of earlier multiscale models The local response function is determined by the stimulus-dependent LGN drive and three cortical input channels: L6 feedback, L4 recurrent excitation, and L4 recurrent inhibition The DNN surrogate approximates each library response function ΨQ by a separately parameterized DNN Ψb Q with 19,297 trainable parameters. Equilibrium computation solves Φ(f) − f = 0, and the stability criterion requires all eigenvalues of L to have negative real parts The response error is measured using the HC-norm, defined as fHC = 0.8(fE)2 + 0.2(fI)2.
Key Findings on Pathway Interaction
The analysis shows that effective L6 feedback and recurrent excitation sustain slowly decaying activity patterns, while recurrent inhibition maintains stability The spatial structure of the strongest recurrent responses closely resembles that of the inhibitory pathway, whereas the slow patterns resemble those supported by L6 feedback Compensation calculations show that coordinated changes in pathway strengths can approximately preserve aggregate E/I firing-rate maps and broad orientation tuning while changing transient amplitudes and slow recovery times. The effects on steady firing rates can cancel without preserving the feedback evoked by a disturbance These complementary pathway contributions to a coupled circuit show that amplification and recovery over time depend on their interaction. p. 8 Together, pooling and pathway cancellation help explain why the recurrent circuit responds
weakly to many population changes. We note that pathway effects cancel for selected activity
perturbations at fixed circuit parameters, whereas the later compensation analysis varies pathway
strengths together to preserve steady responses approximately.Where the paper says it
Experimental Relevance
Experimental tests of recovery and pathway contributions can be examined by modulating inhibitory drive within L4 and the activity of L6 using GABAergic optogenetic manipulation Comparing responses to weak perturbations under a fixed visual stimulus can reveal transient amplification, slow decay, and shifts in a population orientation readout Spatially pooled responses and their recovery trajectories can already help distinguish candidate circuits with similar steady tuning The paper suggests that detailed recordings would require denser sampling distributed across orientation domains to resolve the spatial organization of slow modes.
Computational Scope and Limitations
The computational scope is estimated at order 105 over spiking simulation after library construction and training The added population-rate equation describes the evolution of coarse firing-rate statistics; agreement of steady responses does not by itself validate its transient dynamics The present stability analysis is local to the tested responses and does not exclude other attractors The recovery times also depend on the common time calibration, which is held fixed across pathway comparisons.
Future Directions
A natural extension is to test whether these pathway relationships persist across spatial frequency, temporal frequency, and motion direction as well as orientation and contrast Extending the present analysis across stimulus families could reveal which pathway combinations preserve feature distinctions and which amplify disturbances that obscure them Building and comparing larger mechanistic models will require testing many uncertain parameters and architectures Fast CG evaluation can reduce the cost of each test, while generative methods can amortize parameter inference across target responses through a shared simulation library. The paper suggests that comparing both steady representations and recovery under matched perturbations would then test whether similar visual functions are supported by similar circuit mechanisms. p. 0 Similar visual tuning conceals different recovery dynamics in a model
of macaque V1
Xialei Huang1
, Qingyu Yan2
, Zheng Bian3,4,∗
, and Zhuo-Cheng Xiao1,5,6,∗
1Division of Art and Science, NYU Shanghai, Shanghai 200124, China
2Applied Physics Graduate Program, Rice University, Houston, TX 77005, USA
3
Instituto de Matemática, Estatística e Ciência da Computação da Universidade de São Paulo, São Paulo,
SP 05508-090, Brazil
4Department of Mathematics, Imperial College London, London SW7 2AZ, UK
5NYU-ECNU Institute of Mathematical Sciences at NYU Shanghai, Shanghai 200124, China
6NYU-ECNU Institute of Brain and Cognitive Science at NYU Shanghai, Shanghai 200124, China
∗Corresponding authors
Abstract
Biologically detailed neuronal networks can model how brain circuits perform observed
functions, but their high computational cost limits the comparisons needed to test biological
hypotheses. To address this limitation, we extend a multiscale approach in which responses of
small populations of spiking neurons are computed in advance and reused across a larger cortical
circuit.Where the paper says it
Key Mathematical Relationships
The chain rule applied to the local response networks and input maps gives J = J6 + JE + JI The pathway contribution is defined as ca = Re u∗Jav u∗v, −1 + c6 + cE + cI = Re λ − 1 The recovery time plotted in Fig. 7B is the asymptotic envelope time τ /[1−α(J)], where α(J) = max Re λ(J) The bilinear form is that induced by the E/I response metric in Eq. p. 2 Also, we impose that
the learned functions be smooth and hence provide derivatives: they describe how small changes
in a population’s inputs change its firing rate. Combined with the spatial connections among
populations, these derivatives form the local response matrix, or Jacobian, which describes how a
small perturbation is transmitted across the whole circuit. Studies of other neuronal circuits show that different parameter
combinations can generate similar ongoing activity 12, yet experiments showed that two circuits
producing similar dynamics can respond differently to the same perturbation13. In V1, coordinated
changes in recurrent excitation, inhibition, and feedback from other layers may approximately
preserve a steady firing-rate map while substantially changing transient responses. The connection kernels linearly convolve the input changes
received from other pixels, while the local response slopes specify how much the pooled input rates
change the out firing rates. Combining these steps gives the derivative of the full cortical response. Once trained, the same functions are used at every pixel. A fitted function supplies response values between the sampled inputs. Freezing L6 feedback moves
the leading eigenvalue away from the stability boundary (β6 = −1, Fig. Freezing inhibitory
input has the opposite effect: originally negative growth rates become positive, corresponding to
eigenvalues of J beyond that boundary (βI = −1, Fig. Following the drifting-grating measurements used to characterize macaque
V1 tuning4
, we assess OS by recording the excitatory rate of each model pixel as the grating angle
is varied. The two pixels marked in Fig. Freezing L6 feedback moves
the leading eigenvalue away from the stability boundary (β6 = −1, Fig. Freezing inhibitory
input has the opposite effect: originally negative growth rates become positive, corresponding to
eigenvalues of J beyond that boundary (βI = −1, Fig. (1)
The rate equation supplies a temporal evolution for the population means. Fast spike fluctuations
and gamma-phase dynamics are averaged out in computing the local equilibria and are neglected
as time-resolved dynamics here. 34 ms by aligning the peak time of maximized
transient gain with V1 response timing. For each observation interval, we find the small perturbation
6Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Visual Readout Analysis
The readout in Fig. 5E measures small orientation deviations around a reference stimulus, showing that most of the decoding deviation during recovery comes from preferred and neighboring domains The spatial profiles also distinguish the two pathway-associated patterns (Fig. p. 9 Consistently, most of the decoding deviation during recovery
comes from the preferred and neighboring domains, while the orthogonal-domain contribution is
small. The spatial profiles and readout thus suggest that the tested disturbances mainly distort the
representation near the presented orientation, rather than strongly recruit an orthogonal-orientation
pattern. Consistently, most of the decoding deviation during recovery
comes from the preferred and neighboring domains, while the orthogonal-domain contribution is
small. The spatial profiles and readout thus suggest that the tested disturbances mainly distort the
representation near the presented orientation, rather than strongly recruit an orthogonal-orientation
pattern.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Deep surrogate networks can be used to approximate local response functions, achieving an estimated acceleration of order 105 for steady-state evaluation relative to spiking network simulation by replacing lookup and interpolation with a DNN approximation.
-
The smooth deep surrogate of local responses provides derivatives, allowing for the analysis of linear stability and tracing how a perturbation in brain activity is transmitted through different pathways.
-
The system can analyze pathway contributions to stability, strong recurrent responses, and slow recovery by deriving a population-rate equation and analyzing its linearized dynamics near steady state.
-
The approach allows for comparing pathway combinations that produce similar responses to visual stimuli while producing different recovery dynamics by using fixed-point-preserving interventions to compare pathway contributions at a common steady response.
-
The improved AI system can systematically test biological hypotheses across stimuli and biological hypotheses by offering efficient evaluation with an interpretable circuit structure, making systematic comparisons practical.
Abstract
Biologically detailed neuronal networks can model how brain circuits perform observed functions, but their high computational cost limits the comparisons needed to test biological hypotheses. To address this limitation, we extend a multiscale approach in which responses of small populations of spiking neurons are computed in advance and reused across a larger cortical circuit. In particular, we approximate these local responses by deep neural networks and achieve an estimated acceleration of the order 10 5 for steady-state evaluation relative to spiking network simulation. Beyond computational acceleration, the smooth deep surrogate of local responses provides derivatives, allowing us to analyze linear stability and to trace how a perturbation in brain activity is transmitted through different pathways. We demonstrate the approach in a model of an input layer of macaque primary visual cortex (V1), testing circuits under visual stimulation and controlled perturbations. We resolve pathway contributions to stability, strong recurrent responses, and slow recovery. Yet different combinations of pathway strengths can support similar responses to visual stimuli while producing different recovery dynamics. Therefore, Steady responses and perturbation recovery place complementary constraints on candidate circuits. Overall, our approach offers efficient evaluation with an interpretable circuit structure, making systematic comparisons across stimuli and biological hypotheses practical.
Sources
- Mostly-monocular responses and other visual functions in a multiscale network model of Macaque V1
- Diffusion learning reveals viable parameter manifolds and compensation geometry in biological dynamical systems
Related papers
- Self-Supervised Learning of Local Predictive Directions as a Neural Computation
- Synchronized Hearts, Desynchronized Brains: A Nonlinear Dynamics Perspective for Function and Failure
- Message interaction dynamics covary with brain volume across mammalian connectomes
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
- The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models
- BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks