Self-Supervised Learning of Local Predictive Directions as a Neural Computation
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: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We're looking at the paper: "Self-Supervised Learning of Local Predictive Directions as a Neural Computation".
Marcus: Here's the gist of the paper. Natural sensory dynamics are globally nonlinear but often exhibit simple local geometry, with nearby trajectories evolving preferentially along a small number of predictive directions The paper introduces Local Predictive Directions (LPD), a self-supervised framework that learns local directions whose projections maximally predict near-future trajectory evolution On stochastic nonlinear dynamical systems, LPD recovers the local predictive geometry and partitions trajectories into distinct dynamical regimes On naturalistic sensory streams, the online algorithm produces center–surround spatial receptive fields and ON/OFF sustained and transient temporal filters resembling those of early visual neurons.
Ines: Core Methodology The LPD framework utilizes past–future canonical correlation analysis (CCA) applied to temporal differences to discover local predictive directions Temporal differences, defined as delta xt:= xt+τ − xt, are used because they make the method invariant to translations of the local coordinate system and avoid estimation of equilibria or a global dynamical model.
Marcus: The core objective is a non-negative similarity-matching formulation which partitions trajectory segments into predictive domains, each associated with a distinct local predictive direction.
Ines: This formulation preserves the predictive correlation structure of the data while non-negativity induces competition and locality, causing output units to specialize to different regions of state space.
Title and authors: Ines: So we're looking at this paper called "Self-Supervised Learning of Local Predictive Directions as a Neural Computation." It’s about how neural circuits might figure out the local geometry of natural sensory dynamics directly from experience.
Marcus: Right. We saw that these systems are globally nonlinear, but locally, nearby trajectories follow a few simple directions. This paper introduces something called Local Predictive Directions, or LPD, which learns those specific local directions whose projections predict the near future of the trajectory.
Yuki: From a population genetics perspective, what this suggests is that organisms might be using these local rules—these predictive directions—to navigate their environment without needing a full map of everything happening everywhere.
Ines: Exactly. The core idea is using past and future canonical correlation analysis on temporal differences to find these directions, which makes the method translation invariant because it doesn't care where the system is in its state space, like near an unknown saddle point.
Marcus: And this isn't just about finding one direction; it uses a non-negative similarity matching formulation to partition trajectory segments into different domains, each associated with a distinct local predictive direction. This forces the output units to specialize to different regions of the state space because of that non-negativity constraint inducing competition and locality.
Yuki: That sounds like how an animal might rapidly switch behaviors based on what it's seeing right now, reacting to immediate local patterns rather than thinking about long-term global states.
Ines: And when we look at the online implementation, it turns this into a neural network interpretation where the feedforward weights correspond to local predictive filters over past increments, and the lateral weights implement competition between units. The objective function they minimize is structured to find those directions while respecting that non-negativity constraint.
Marcus: The validation on dynamical systems was pretty solid; they showed it correctly identifies unstable directions in affine saddle point dynamics even when the exact location of that fixed point wasn't known, and for generalized Lotka volterra systems, it captured the direction associated with each transition between saddles in a stable heteroclinic channel.
Title and authors: Yuki: That’s interesting because those transitions between different dynamical regimes are often what drive population structure or survival in nature. So this method is finding the underlying mechanism of those regime switches in a data-driven way.
Ines: And for sensory data, the results were quite striking; when applied to naturalistic stimuli, LPD learned temporal filters that look like electrophysiologically measured neuronal receptive fields, including center surrounded structures and sustained and transient temporal filters resembling early visual neurons.
Marcus: That means the framework suggests that sensory neurons might actually be detectors of these local predictive directions governing the stimulus dynamics. It’s a self-supervised mechanism where the neural circuit discovers the local geometry of its environment without needing any labeled data to start.
Yuki: If we think about how we study evolution, this hints that perhaps the initial structures in sensory systems are tuned precisely to capture these simple, local dynamics that govern immediate survival.
Ines: And they even have a criterion for selecting the number of output components k based on where the objective function plateaus. They say this elbow in the objective curve provides a natural estimate of how many dynamical modes there actually are, coinciding with the precision and recall optima for those learned directions.
Marcus: They also compared it to Gaussian mixture models, and found that LPD captures more predictive information at small lags than GMM because it learns the magnitude of those finite differences rather than being constrained by normalization conditions. It also showed that both finite differences and non-negativity are necessary for optimal performance, converging with finite differences only at high noise levels.
Yuki: That robustness across different noise levels is important; in a real biological setting, the signal is never perfectly clean. So this method seems well-suited for messy real-world data streams.
Ines: And they showed that the offline and online implementations give equivalent results when the underlying dynamical process is stationary and the dataset is large enough, which makes it practical for deployment. The learned local predictive directions correspond to modes where small perturbations influence future trajectories, suggesting a natural interface with control systems.
Marcus: So to summarize this paper on "Self-Supervised Learning of Local Predictive Directions as a Neural Computation," they propose LPD to discover the local geometry of time series by learning directions that predict the near future using past and future canonical correlation analysis on temporal differences.
Title and authors: Yuki: It sounds like a way to decode the immediate, local rules governing complex systems without needing to know the global dynamics beforehand.
Ines: And it shows this process can manifest in neural circuits as filters resembling sensory receptive fields, which is a really powerful way to connect abstract mathematical concepts to biology.
Marcus: The key improvement here is using temporal differences instead of absolute states, which makes it invariant to translations of the local coordinate system and avoids having to estimate equilibria or build a global dynamical model.
Yuki: That invariance is key when you're trying to study things in a dynamic environment where you don't know the exact starting conditions.
Ines: And they give us a way to select the number of directions we need by watching where the learning objective flattens out, which gives us an estimate of how many underlying dynamical modes are actually present.
Marcus: For anyone working with large cohorts or noisy time series, this method offers a way to find those local predictive regimes and partition the data into distinct dynamical domains.
Yuki: It’s fascinating to think about how this idea applies beyond dynamics, perhaps even to understanding the structure of genetic variation in populations based on immediate environmental pressures.
Ines: And we have to remember their limitation; they mention that performance is only weakly affected by the choice of the time lag tau, as long as it stays below the characteristic timescales of the system dynamics. They also note that convergence with finite differences only happens at high noise levels, meaning both components are important for optimal results.
Marcus: So, in short, LPD provides a self-supervised framework to learn local tangent space directions from trajectories, which has shown success in identifying unstable manifolds and generating filters that mimic sensory neurons.
Yuki: It’s a solid piece of work showing how simple local geometry can be extracted from very complex systems through clever statistical matching.
Ines: It certainly gives us a new lens for looking at how biological systems might encode their immediate predictive capabilities into their fundamental structure.
The paper's summary: Ines: So, we're talking about how this new framework, Local Predictive Directions or LPD, tries to learn the local geometry of natural sensory dynamics by finding directions that predict what happens next on a trajectory.
Marcus: Yeah, it’s all about using past and future data to find these specific directions that are invariant to where the system is in its state space while avoiding having to estimate a whole global dynamical model.
Yuki: From a population genetics angle, what this suggests is that organisms might be using these local rules—these predictive directions—to navigate their environment without needing a full map of everything happening everywhere.
Ines: Exactly. The core idea here is that instead of looking at the whole system, you look at how nearby trajectories move together locally over time, and you use past and future data to find those specific directions that project that near-future movement.
Marcus: And they achieve this by using past and future canonical correlation analysis on temporal differences, which makes the method translation invariant because it doesn't care about the exact starting conditions or where the system is in its state space.
Yuki: That invariance is important when you're trying to study things in a dynamic environment where you don't know the exact initial setup, like how a species might react to sudden changes in habitat.
Ines: And they use this non-negative similarity matching formulation to partition trajectory segments into different domains, each with its own local predictive direction, which forces the output units to specialize based on where they are in that space.
Marcus: That’s the mechanism for finding those distinct regimes; it uses a constraint that makes the units compete locally, pushing them toward different areas of the state space.
Yuki: So this isn't just finding one general rule for a species; it's finding many local rules that govern how that species reacts in specific patches of its environment.
Ines: And when you look at the online version, it turns this into a neural network structure where the weight matrix defines local predictive filters over past increments, and the lateral matrix implements competition between those units.
Marcus: The validation on dynamical systems was pretty strong; they showed it correctly identifies unstable directions in affine saddle point dynamics even when the exact location of that fixed point wasn't known.
Yuki: That’s interesting because those transitions between different dynamical regimes are often what drive population structure or survival in nature, so finding them locally is a big deal.
Ines: And for sensory data, they applied it to naturalistic stimuli and found temporal filters that look qualitatively like electrophysiologically measured neuronal receptive fields, including center surrounded structures and sustained and transient filters.
Marcus: That means the framework suggests that sensory neurons might actually be detectors of these local predictive directions governing stimulus dynamics, a really cool connection.
Yuki: If we think about how we study evolution, this hints that perhaps the initial structures in sensory systems are tuned precisely to capture these simple, local dynamics that govern immediate survival.
Ines: And they give us a criterion for selecting the number of output components based on where the learning objective plateaus, which provides a natural estimate of how many dynamical modes there actually are.
Marcus: That elbow in the objective curve is meant to be an internal way to select the number of directions, aligning with precision and recall optima for those learned directions.
Yuki: So this method gives us a way to estimate complexity without needing an external guess, which is useful when you're trying to model something as rich as a biological brain or a complex ecosystem.
Ines: And they also compared it to Gaussian mixture models and found that LPD captures more predictive information at small lags because it learns the magnitude of those finite differences directly.
Marcus: But they also noted that both using finite differences and using non-negativity constraints are essential for optimal performance, converging with finite differences only at high noise levels.
Yuki: That robustness across different noise levels is important; in a real biological setting, the signal is never perfectly clean, so this method seems well-suited for messy real-world data streams.
Ines: And they showed that the offline and online implementations give equivalent results when the underlying dynamical process is stationary and the dataset is large enough, making it practical for deployment.
Marcus: The learned local predictive directions correspond to modes where small perturbations influence future trajectories, which suggests a natural interface with control systems.
Yuki: It’s fascinating to think about how this idea applies beyond dynamics, perhaps even to understanding the structure of genetic variation in populations based on immediate environmental pressures.
Ines: So this paper gives us a self-supervised way to learn local geometry from trajectories, and it shows that these learned directions can appear as filters in neural circuits.
Marcus: It’s a solid piece of work showing how simple local geometry can be extracted from very complex systems through careful statistical matching.
Yuki: It certainly gives us a new lens for looking at how biological systems might encode their immediate predictive capabilities into their fundamental structure.
The paper's improvements: Tom: So, we’re looking at how this paper, Local Predictive Directions or LPD, tries to learn the local geometry of natural sensory dynamics by finding directions that predict what happens next on a trajectory.
Ines: It basically says that instead of modeling the whole system globally, you can use past and future data to find these specific directions that project near-future movement.
Marcus: Yeah, it uses past and future canonical correlation analysis on temporal differences to find those directions that are invariant to where the system is in its state space.
Yuki: That invariance is important when you're trying to study things in a dynamic environment where you don't know the exact initial setup, like how a species might react to sudden changes in habitat.
Ines: They use this non-negative similarity matching formulation to partition trajectory segments into different domains, each with its own local predictive direction, which forces the output units to specialize based on where they are.
Marcus: That’s the mechanism for finding those distinct regimes; it uses a constraint that makes the units compete locally, pushing them toward different areas of the state space.
Yuki: So this isn't just finding one general rule for a species; it's finding many local rules that govern how that species reacts in specific patches of its environment.
Ines: And when you look at the online version, it turns this into a neural network structure where the weight matrix defines local predictive filters over past increments, and the lateral matrix implements competition between those units.
Marcus: The validation on dynamical systems was pretty strong; they showed it correctly identifies unstable directions in affine saddle point dynamics even when the exact location of that fixed point wasn't known.
Yuki: That’s interesting because those transitions between different dynamical regimes are often what drive population structure or survival in nature, so finding them locally is a big deal.
Ines: And for sensory data, they applied it to naturalistic stimuli and found temporal filters that look qualitatively like electrophysiologically measured neuronal receptive fields.
Marcus: That means the framework suggests that sensory neurons might actually be detectors of these local predictive directions governing stimulus dynamics, a really cool connection.
Yuki: If we think about how we study evolution, this hints that perhaps the initial structures in sensory systems are tuned precisely to capture these simple, local dynamics that govern immediate survival.
Ines: And they give us a criterion for selecting the number of output components based on where the learning objective plateaus, which provides a natural estimate of how many dynamical modes there actually are.
Marcus: That elbow in the objective curve is meant to be an internal way to select the number of directions, aligning with precision and recall optima for those learned directions.
Yuki: So this method gives us a way to estimate complexity without needing an external guess, which is useful when you're trying to model something as rich as a biological brain or a complex ecosystem.
Ines: And they also compared it to Gaussian mixture models and found that LPD captures more predictive information at small lags because it learns the magnitude of those finite differences directly.
Marcus: But they also noted that both using finite differences and using non-negativity constraints are essential for optimal performance, converging with finite differences only at high noise levels.
Yuki: That robustness across different noise levels is important; in a real biological setting, the signal is never perfectly clean, so this method seems well-suited for messy real-world data streams.
Ines: And they showed that the offline and online implementations give equivalent results when the underlying dynamical process is stationary and the dataset is large enough, making it practical for deployment.
Marcus: The learned local predictive directions correspond to modes where small perturbations influence future trajectories, which suggests a natural interface with control systems.
Yuki: It’s fascinating to think about how this idea applies beyond dynamics, perhaps even to understanding the structure of genetic variation in populations based on immediate environmental pressures.
Conclusion: Ines: So, to wrap up, this paper on "Self-Supervised Learning of Local Predictive Directions as a Neural Computation" shows that AI can learn local geometry by finding directions that predict near-future movement using past and future data.
Marcus: It really highlights how you can bypass estimating a whole global dynamical model by focusing only on those small, local predictive directions.
Yuki: From my perspective, it suggests that even in complex systems, the immediate rules governing what happens next are often the most important for survival and structure at a population level.
Ines: And the result is that these learned directions can look just like filters in real neurons, which means we might be looking at a direct computational interpretation of sensory processing.
Marcus: The methodology uses temporal differences to make everything translation invariant, which is smart because it lets us ignore where the system starts in its state space.
Yuki: But I have to ask what this sample really supports; does finding these local rules in a lab setting translate well to the massive variation you see across a real population?
Ines: The paper’s validation on stochastic systems and sensory data suggests it works well for those transitions, but they did flag that performance only converges with finite differences at high noise levels.
Marcus: So, while the method is robust, it definitely needs both the finite difference part and the non-negativity constraint to really hit its best performance in noisy data.
Yuki: It’s interesting because if we can find these local rules reliably, maybe we can better understand why certain genetic variants persist or how populations adapt quickly.
Ines: And their internal criterion for selecting the number of output components based on where the learning objective plateaus is a useful way to estimate the actual number of dynamical modes present.
Marcus: That’s a solid piece of engineering, using the objective function itself to guide how many parameters you need.
Yuki: It gives us a concrete way to measure complexity that isn't just guessing based on external criteria.
Ines: So, the implication is that AI can discover these local rules in environments—whether it’s fluid dynamics or sensory streams—without needing any prior knowledge of the underlying physics.
Marcus: It opens up a new path for modeling systems where you don't have enough data to build a big picture model, but you can still capture the important local interactions.
Yuki: This work on "Self-Supervised Learning of Local Predictive Directions as a Neural Computation" is really showing us that the most complex behaviors often come from simple, local dynamics that we just need to discover.
Marco Zenari, Xuehao Ding, Daniel Burnham, Austin R. Ellis-Mohr, Anirvan M. Sengupta, Dmitri B. Chklovskii
University of Padova · Flatiron Institute, New York, NY, USA · University of Oregon, Eugene, OR, USA · University of Illinois at Urbana-Champaign · Rutgers University, New Brunswick, NJ, USA
q-bio.NC
Submitted: 2026-11-05
Updated: 2026-11-05
Comments: Marco Zenari and Xuehao Ding contributed equally
Code: https://github.com/MarcoZenari/local-predictive-directions
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: Natural sensory dynamics are globally nonlinear but often exhibit simple local geometry, with nearby trajectories evolving preferentially along a small number of predictive directions The paper…
Key concepts
- Local Predictive Directions (LPD)
- LPD discovers small, local directions in the state space along which nearby trajectories evolve predictably. It is learned by finding directions that best predict future trajectory changes from past increments, allowing the system to capture complex dynamics locally.
- Temporal Differences
- Instead of using absolute states, LPD uses temporal differences (xt+τ - xt) to define descriptors. This makes the learning process invariant to shifts in the coordinate system and avoids needing to estimate a global dynamical model or fixed points.
- Non-negative Similarity-Matching Objective
- The core objective function partitions trajectory segments into distinct predictive domains using a non-negative formulation. This constraint induces competition between units, causing the learned components to specialize in different regions of the state space, effectively mapping out local dynamical regimes.
Terminology
Summary
Natural sensory dynamics are globally nonlinear but often exhibit simple local geometry, with nearby trajectories evolving preferentially along a small number of predictive directions The paper introduces Local Predictive Directions (LPD), a self-supervised framework that learns local directions whose projections maximally predict near-future trajectory evolution On stochastic nonlinear dynamical systems, LPD recovers the local predictive geometry and partitions trajectories into distinct dynamical regimes On naturalistic sensory streams, the online algorithm produces center–surround spatial receptive fields and ON/OFF sustained and transient temporal filters resembling those of early visual neurons. p. 0 On stochastic nonlinear dynamical systems, LPD recovers the local predictive geometry and
partitions trajectories into distinct dynamical regimes. On naturalistic sensory streams, the
online algorithm produces center–surround spatial receptive fields and ON/OFF sustained
and transient temporal filters resembling those of early visual neurons.Where the paper says it
Core Methodology
The LPD framework utilizes past–future canonical correlation analysis (CCA) applied to temporal differences to discover local predictive directions Temporal differences, defined as delta xt:= xt+τ − xt, are used because they make the method invariant to translations of the local coordinate system and avoid estimation of equilibria or a global dynamical model. The core objective is a non-negative similarity-matching formulation which partitions trajectory segments into predictive domains, each associated with a distinct local predictive direction. p. 0 LPD applies past–future canonical correlation analysis to temporal differences,
making the learned representation invariant to translations of the underlying state while
avoiding estimation of equilibria or a global dynamical model. To represent systems with
multiple predictive regimes, we introduce a non-negative similarity-matching objective,
with offline and online algorithms, that partitions trajectory segments into predictive domains, each associated with a distinct local predictive direction. Here,
similarity matching preserves the predictive correlation structure of the data while nonnegativity induces competition and locality, Fig. 1b; output units therefore specialize to
different regions of state space, learning distinct local predictive directions, Fig.Where the paper says it
Where the paper says it
Learning Mechanism
The online algorithm provides a natural neural network interpretation where feedforward weights W correspond to local predictive filters over past increments and lateral weights M implement competition between units. The objective (7), min Z≥0 GG⊤ − Z⊤Z F, is solved via an optimization of Eq. p. 5 min
Z≥0
min
N
max
M
Tr h
− 4Z
⊤NG⊤ + 2Z
⊤MZ + 2N⊤N − M⊤M
i
. (8)
Optimization of Eq. (8)
Optimization of Eq. (8) using gradient descent alternates an inference step for Z at fixed
W,M with updates of W and M from empirical covariances, where the weight matrix W
is defined by WP = NG⊤; the resulting batch algorithm is given in Appendix H. Rectification enforces nonnegative activity, so each unit reports the presence of a particular local predictive direction
rather than a signed canonical coordinate. The resulting activity vector zt therefore acts
as a soft domain assignment of the current trajectory segment to one of several locally
predictive tangent-space regimes (Fig.Where the paper says it
Where the paper says it
Where the paper says it
Validation on Dynamical Systems
The LPD method was validated on stochastic nonlinear systems with known local structure, such as affine saddle-point dynamics, where it correctly identifies the unstable directions even when the fixed-point location is unknown. For generalized Lotka–Volterra (gLV) systems, LPD successfully captures the predictive direction associated with each transition between saddles in a stable heteroclinic channel. p. 6 To test multiple saddle points, we apply LPD
to a stochastic generalized Lotka–Volterra (gLV) system, widely used to study stable heteroclinic channels (SHC) (Bick and Rabinovich, 2010; Afraimovich et al. We consider
N = 3 species arranged to form an SHC, producing sequential transitions between saddles at
which one species dominates and the others are nearly extinct (Appendix I).Where the paper says it
Validation on Sensory Data
When applied to naturalistic stimuli, LPD learned temporal filters qualitatively resemble electrophysiologically measured neuronal receptive fields, including center–surround structures and ON/OFF sustained and transient filters. These learned filters are interpreted as detectors of LPDs governing stimulus dynamics, suggesting that sensory neurons can be viewed as detectors of these local predictive directions. p. 8 tions are nearly isotropic in space, whereas time has a fixed arrow. These results suggest
that sensory neurons can be viewed as detectors of LPDs governing stimulus dynamics. On naturalistic sensory streams, the
online algorithm produces center–surround spatial receptive fields and ON/OFF sustained
and transient temporal filters resembling those of early visual neurons. Our results suggest
that predictive learning can provide a self-supervised mechanism by which neural circuits
discover the local geometry of environmental dynamics and represent it in receptive fields.Where the paper says it
Where the paper says it
Model Selection Criterion
The proposed normative framework offers an internal, task-aligned criterion for selecting the number of output components k by observing where the objective function plateaus, which provides a natural estimate of the number of dynamical modes. This elbow in the objective curve coincides with the precision and recall optima of the learned directions. p. 7 Zenari* Ding* Burnham Ellis-Mohr Sengupta Chklovskii
clustering or Hidden Markov Models: the objective plateaus once k reaches the intrinsic
dimensionality of the predictive structure, and this heuristic elbow coincides with the precision and recall optima of the learned directions (Appendix J; Appendix J. 1 for natural
stimuli). Zenari* Ding* Burnham Ellis-Mohr Sengupta Chklovskii
clustering or Hidden Markov Models: the objective plateaus once k reaches the intrinsic
dimensionality of the predictive structure, and this heuristic elbow coincides with the precision and recall optima of the learned directions (Appendix J; Appendix J. 1 for natural
stimuli). Zenari* Ding* Burnham Ellis-Mohr Sengupta Chklovskii
clustering or Hidden Markov Models: the objective plateaus once k reaches the intrinsic
dimensionality of the predictive structure, and this heuristic elbow coincides with the precision and recall optima of the learned directions (Appendix J; Appendix J. 1 for natural
stimuli).Where the paper says it
Where the paper says it
Where the paper says it
Robustness and Comparison
LPD combines finite differences for translational invariance and a non-negativity constraint, which are both essential for optimal performance compared to standard past–future CCA or unconstrained finite-differences CCA. The results show that LPD achieves higher NMI at low and intermediate noise, converging with FD only at high noise, indicating that both ingredients matter. p. 7 2b) against past–future CCA (PF), non-negative PF (NN-PF)
and finite-differences CCA (FD), scoring held-out trajectories by the Normalized Mutual
Information (NMI) between the maximally active component and the next saddle visited. At low and intermediate noise LPD attains higher NMI than all three, converging with
FD only at high noise, so both ingredients matter (Appendix L). At low and intermediate noise LPD attains higher NMI than all three, converging with
FD only at high noise, so both ingredients matter (Appendix L). LPD also carries more
information about the future tangent vector than a Gaussian Mixture Model fit to the same
finite differences, partly via its unnormalized magnitude (Appendix M). , 2009), the proposed objective clusters trajectory segments by short-time predictive structure. The resulting algorithm admits offline
and online implementations, the online one having a natural neural interpretation in terms
of feedforward weights, lateral competition, and rectified activity. An important direction for future work is to establish theoretical conditions under which non-negative similarity matching provably recovers
LPDs. Finally, the learned LPDs correspond to modes where small perturbations influence
future trajectories, suggesting a natural interface with control.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Code Availability
Code to reproduce all experiments is available at https://github.com/MarcoZenari/local-predictive-directions. p. 0 Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. arXiv:2611. 07809v1 [q-bio. arXiv:2611. 07809v1 [q-bio. arXiv:2611. 07809v1 [q-bio. arXiv:2611. 07809v1 [q-bio. 07809v1 [q-bio. NC] 5 Nov 2026 07809v1 [q-bio. NC] 5 Nov 2026 07809v1 [q-bio. NC] 5 Nov 2026 07809v1 [q-bio. NC] 5 Nov 2026 Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. edu
Rutgers University, New Brunswick, NJ, USA
Flatiron Institute, New York, NY, USA
Dmitri B. Chklovskii
dchklovskii@flatironinstitute. edu
Rutgers University, New Brunswick, NJ, USA
Flatiron Institute, New York, NY, USA
Dmitri B. Chklovskii
dchklovskii@flatironinstitute. Zenari* Ding* Burnham Ellis-Mohr Sengupta Chklovskii
1. Introduction
High-dimensional time series from physical, biological, and neural systems often evolve on
structured nonlinear manifolds. Zenari* Ding* Burnham Ellis-Mohr Sengupta Chklovskii
1. Introduction
High-dimensional time series from physical, biological, and neural systems often evolve on
structured nonlinear manifolds. Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. org
Flatiron Institute, New York, NY, USA
Daniel Burnham
dburnham@uoregon. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. org
Flatiron Institute, New York, NY, USA
Daniel Burnham
dburnham@uoregon. edu
University of Oregon, Eugene, OR, USA
Austin R. Ellis-Mohr
austine4@illinois. edu
University of Oregon, Eugene, OR, USA
Austin R. Ellis-Mohr
austine4@illinois. edu
University of Illinois at Urbana-Champaign,
Urbana, IL, USA
Anirvan M. Sengupta
anirvans@physics. edu
University of Illinois at Urbana-Champaign,
Urbana, IL, USA
Anirvan M. Sengupta
anirvans@physics. Chklovskii. arXiv:2611. Chklovskii. arXiv:2611. Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute. Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations
Self-Supervised Learning of Local Predictive Directions as a
Neural Computation
Marco Zenari*
marco. it
University of Padova, Padova, Italy
Xuehao Ding*
xding@flatironinstitute.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
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
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
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
References
George A Ackerson and King-Sun Fu. On state estimation in switching environments. p. 9 Zenari* Ding* Burnham Ellis-Mohr Sengupta Chklovskii
References
George A Ackerson and King-Sun Fu. On state estimation in switching environments. IEEE
Transactions on Automatic Control, 15(1):10–17, 1970. Valentin Afraimovich, Irma Tristan, Ramon Huerta, and Mikhail I Rabinovich. Valentin Afraimovich, Irma Tristan, Ramon Huerta, and Mikhail I Rabinovich. Winnerless
competition principle and prediction of the transient dynamics in a lotka–volterra model. Chaos: An Interdisciplinary Journal of Nonlinear Science, 18(4), 2008. KS Arun and SY Kung. KS Arun and SY Kung. Balanced approximation of stochastic systems. SIAM journal on
matrix analysis and applications, 11(1):42–68, 1990. Horace B Barlow. Horace B Barlow. Summation and inhibition in the frog’s retina. The Journal of physiology,
119(1):69, 1953. George Keith Batchelor. George Keith Batchelor. An introduction to fluid dynamics. Cambridge university press,
2000. William Bialek, Ilya Nemenman, and Naftali Tishby. William Bialek, Ilya Nemenman, and Naftali Tishby. Predictability, complexity, and learning. Neural Computation, 13(11):2409–2463, 2001. Christian Bick and Mikhail I Rabinovich. Christian Bick and Mikhail I Rabinovich. On the occurrence of stable heteroclinic channels
in lotka–volterra models. Dynamical Systems, 25(1):97–110, 2010. Steven L Brunton, Marko Budiˇsi´c, Eurika Kaiser, and J Nathan Kutz. Steven L Brunton, Marko Budiˇsi´c, Eurika Kaiser, and J Nathan Kutz. Modern koopman
theory for dynamical systems. Modern koopman
theory for dynamical systems. arXiv preprint arXiv:2102. 12086, 2021. Martin Casdagli. Martin Casdagli. Chaos and deterministic versus stochastic non-linear modelling. Journal
of the Royal Statistical Society: Series B, 54(2):303–328, 1991. Matthew Chalk, Olivier Marre, and Gaˇsper Tkaˇcik. Matthew Chalk, Olivier Marre, and Gaˇsper Tkaˇcik. Toward a unified theory of efficient,
predictive, and sparse coding. Proceedings of the national academy of sciences, 115(1):
186–191, 2018. Gal Chechik, Amir Globerson, Naftali Tishby, and Yair Weiss. Past–future correlation offers a principled definition of such predictive directions. Canonical correlation analysis (CCA) finds projections of paired variables with maximal correlation (Hotelling, 1936); related past–future covariance methods underlie subspace system
identification and balanced stochastic realization (Arun and Kung, 1990; Katayama, 2005),
and connect to predictive information and the Gaussian information bottleneck (Bialek
et al. Journal of Machine Learning Research, 6:165–188, 2005. BG Cleland, WR Levick, and KJ Sanderson. BG Cleland, WR Levick, and KJ Sanderson. Properties of sustained and transient ganglion
cells in the cat retina. The Journal of physiology, 228(3):649–680, 1973. Predrag Cvitanovi´c and Domenico Lippolis. Predrag Cvitanovi´c and Domenico Lippolis. Knowing when to stop: How noise frees us
from determinism. Knowing when to stop: How noise frees us
from determinism. In AIP Conference Proceedings, volume 1468, pages 82–126. In AIP Conference Proceedings, volume 1468, pages 82–126. American
Institute of Physics, 2012. American
Institute of Physics, 2012. Doyne Farmer and John J. Global dynamics may be complex, yet short-time evolution
is typically low-dimensional: within a small neighborhood, state changes are captured by a
few tangent directions — the local-linear view underlying classical nonlinear dynamical systems theory (Strogatz, 1994), Koopman and dynamic mode decomposition (Schmid, 2010;
Brunton et al. , 2021), and local models of nonlinear time series (Farmer and Sidorowich,
1987; Casdagli, 1991). Sidorowich. Predicting chaotic time series. Physical Review
Letters, 59(8):845–848, 1987. Emily B Fox, Erik B Sudderth, Michael I Jordan, and Alan S Willsky. Emily B Fox, Erik B Sudderth, Michael I Jordan, and Alan S Willsky. Nonparametric
bayesian learning of switching linear dynamical systems.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
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
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
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
Where the paper says it
Appendix A
Our method is related to existing approaches for modeling dynamical systems, but differs in key aspects. Switching Linear Dynamical Systems (SLDS) model nonlinear dynamics as collections of locally linear regimes indexed by a latent discrete state that evolves according to a Markov chain. p. 1 Introduction
High-dimensional time series from physical, biological, and neural systems often evolve on
structured nonlinear manifolds. Global dynamics may be complex, yet short-time evolution
is typically low-dimensional: within a small neighborhood, state changes are captured by a
few tangent directions — the local-linear view underlying classical nonlinear dynamical systems theory (Strogatz, 1994), Koopman and dynamic mode decomposition (Schmid, 2010;
Brunton et al. Introduction
High-dimensional time series from physical, biological, and neural systems often evolve on
structured nonlinear manifolds. Global dynamics may be complex, yet short-time evolution
is typically low-dimensional: within a small neighborhood, state changes are captured by a
few tangent directions — the local-linear view underlying classical nonlinear dynamical systems theory (Strogatz, 1994), Koopman and dynamic mode decomposition (Schmid, 2010;
Brunton et al. (d) Soft domain
assignments over time for the trajectory in (a). We introduce a self-supervised objective and algorithm for learning such local predictive directions (LPDs) directly from trajectories, Fig.Where the paper says it
Where the paper says it
Where the paper says it
Appendix B
Under the linearized dynamics, vi ∝ l1, Cf fu1 ∝ r1 Thus, the top predictive mode is orthogonal to every eigen-direction except the most unstable direction.
Appendix C
When k = n is the rank of G, we can eigen-decompose G as follows Thus, we have A⊤ = VRU−1 = VRU⊤Cf f, (15).
Appendix D
Temporal differencing removes the translation, while whitening by Cpp and Cf f removes the remaining linear part, leaving the outputs invariant and the filters equivariant. The derivation uses only C′αβ = BCαβB⊤, which holds for the empirical covariances computed from a single trajectory as well as in the population limit, so the invariance is exact at finite sample.
Appendix E
The solution Z∗ obtained from the unconstrained problem Z∗ = arg min GG⊤ − Z⊤Z F is (26) Z∗ = M−1NG⊤ = M−1NF⊤C−1f f Cf pC−1pp P = M−1WP ≃ V⊤P, where ≃ denotes equality up to an orthogonal transformation and component scaling.
Appendix F
The solution Z∗ obtained from the unconstrained problem Z∗ = arg min GG⊤ − Z⊤Z F is (26) Z∗ = M−1NG⊤ = M−1NF⊤C−1f f Cf pC-1pp P = M-1WP ≃ V⊤P, where ≃ denotes equality up to an orthogonal transformation and component scaling.
Appendix G
We define the second-order statistics Cpp = E[ptp⊤t], Cf f = E[ftf⊤t], and the cross-covariances Cpf = E[ptf⊤t], Cf p = C⊤pf. We now consider the matrix G = P⊤C−1pp CpfC−1f f F, 18. p. 4 , uk]
contain the CCA directions. Define the predictive similarity matrix
G := P
⊤C−1
pp CpfC
−1
f f F = Z
⊤
p RZf , (5)
where Rij = ρiδij is the diagonal matrix of canonical correlations and the second equality
is derived in Appendix C for the full-rank case k = n, becoming a rank-k approximation
for k < n.Where the paper says it
Appendix H
The offline implementation of LPD solves the min–max problem of Eq. (8) by alternating gradient updates. p. 5 (8)
Optimization of Eq. (8) using gradient descent alternates an inference step for Z at fixed
W,M with updates of W and M from empirical covariances, where the weight matrix W
is defined by WP = NG⊤; the resulting batch algorithm is given in Appendix H. C+ := (C+ϵI)
−1 denotes
a Tikhonov-regularized inverse and [·]+ elementwise rectification. Initialize W ∈ R
k×n and M ∈ R
k×k with M ≻ 0;
Initialize covariance estimates Cpp, Cf f , Cpf , Cf p, Czp, Czz;
for t ← 1 to T do
Receive increment descriptors (pt
,ft);
Update Cpp, Cf f , Cpf , Cf p;
Inference: update zt for fixed W,M;
repeat
zt ← [zt + ηZ(Wpt − Mzt)]+;
until inference convergence;
Update Czp and Czz;
Update W ← CzpC+
ppCpfC
+
f fCf pC+
pp;
Update M ← Czz;
end
et al.Where the paper says it
Where the paper says it
Algorithm 2
The input is past and future increment matrices P, F; number of output units k. The process involves an inference step to update Z for fixed W,M followed by a learning step that updates W and M from empirical covariances. p. 5 (8)
Optimization of Eq. (8) using gradient descent alternates an inference step for Z at fixed
W,M with updates of W and M from empirical covariances, where the weight matrix W
is defined by WP = NG⊤; the resulting batch algorithm is given in Appendix H.Where the paper says it
Appendix I
The dynamics are given by xt+1 = x∗ + A (xt − x∗) + σ ξt, ξt ∼ N (0, I), where A = diag(λu, λs) is a diagonal matrix with one unstable eigenvalue λu > 1 and one stable eigenvalue 0 < λs < 1.
Appendix J
The proposed normative framework provides an internal, task-aligned criterion for selecting the number of output components k, evaluated using the same objective optimized during learning rather than an external criterion such as k-means clustering or hidden Markov models. The elbow provides a natural estimate of the number of dynamical modes. p. 8 Finally, the learned LPDs correspond to modes where small perturbations influence
future trajectories, suggesting a natural interface with control. Low-rank controllers acting
along these directions may enable efficient steering between metastable regions, analogous
to control along unstable manifolds in classical dynamical systems. Zenari* Ding* Burnham Ellis-Mohr Sengupta Chklovskii
clustering or Hidden Markov Models: the objective plateaus once k reaches the intrinsic
dimensionality of the predictive structure, and this heuristic elbow coincides with the precision and recall optima of the learned directions (Appendix J; Appendix J. 1 for natural
stimuli).Where the paper says it
Where the paper says it
Appendix K
Performance is only weakly affected by the choice of τ, as long as it remains below the characteristic timescales of the system dynamics. The results show that LPD is robust to moderate changes in temporal discretization. p. 6 The
results of Algorithm 1 show that each of the k = 3 output units successfully captures the
predictive direction associated with one transition between saddles, thereby predicting the
identity of the next dominant species. These findings remain robust for a greater number
of species (N ≤ 100) and across different noise levels (see Appendix I.Where the paper says it
Appendix L
LPD achieves higher NMI at low and intermediate noise, converging with FD only at high noise, indicating that both finite differences and non-negativity are essential for optimal performance. The paper's arXiv ID is 2611.07809v1 [q-bio.NC] 5 Nov 2026. The authors are Marco Zenari, Xuehao Ding, Daniel Burnham, Austin R. Ellis-Mohr, Anirvan M Sengupta & Dmitri B Chklovskii. The paper is published in Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations. The copyright is © 2026 M. Zenari>, X. Ding>, D. Burnham, A.R. Ellis-Mohr, A.M. Sengupta & D.B Chklovskii. The arXiv identifier is arXiv:2611.07809v1 [q-bio.NC] 5 Nov 2026. The paper is published in Proceedings of Machine Learning Research 1–28, 2026 Symmetry and Geometry in Neural Representations.
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
-
The AI system can discover local tangent space directions that vary across state space by applying past-future canonical correlation analysis to temporal differences, making it invariant to translations of the underlying state while avoiding estimation of equilibria or a global dynamical model.
-
The system can represent systems with multiple predictive regimes using a non-negative similarity-matching objective, which partitions trajectory segments into predictive domains, each associated with a distinct local predictive direction.
-
The online implementation allows the neural network to produce soft domain assignments, where the weight matrix W corresponds to feedforward synaptic weights that define local predictive filters over past increments, and the lateral matrix M corresponds to lateral interactions implementing competition between units.
-
The system can identify unstable directions in affine saddle-point dynamics even when the fixed-point location is unknown by using temporal differences to define past and future descriptors, which are translation invariant.
-
The system can analyze naturalistic sensory streams to produce center-surround spatial receptive fields and ON/OFF sustained and transient temporal filters resembling those of early visual neurons by applying the LPD algorithm to translated models of natural images.
Abstract
Natural sensory dynamics are globally nonlinear but often exhibit simple local geometry, with nearby trajectories evolving preferentially along a small number of predictive directions. We ask how neural circuits could discover this geometry directly from temporal experience. We introduce Local Predictive Directions (LPD), a self-supervised framework that learns local directions whose projections maximally predict near-future trajectory evolution. LPD applies past-future canonical correlation analysis to temporal differences, making the learned representation invariant to translations of the underlying state while avoiding estimation of equilibria or a global dynamical model. To represent systems with multiple predictive regimes, we introduce a non-negative similarity-matching objective, with offline and online algorithms, that partitions trajectory segments into predictive domains, each associated with a distinct local predictive direction. The online algorithm admits a natural neural-network interpretation: the resulting activities provide soft domain assignments, while synaptic weights encode the corresponding local predictive filters. On stochastic nonlinear dynamical systems, LPD recovers the local predictive geometry and partitions trajectories into distinct dynamical regimes. On naturalistic sensory streams, the online algorithm produces center-surround spatial receptive fields and ON/OFF sustained and transient temporal filters resembling those of early visual neurons. Our results suggest that predictive learning can provide a self-supervised mechanism by which neural circuits discover the local geometry of environmental dynamics and represent it in receptive fields.
Sources
- Modern Koopman Theory for Dynamical Systems
- Representation Learning with Contrastive Predictive Coding
Related papers
- 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
- CTM-AI: Evolving Multi-Agent Intelligence Inspired by a Model of Consciousness