Pith. sign in

REVIEW 4 major objections 5 minor 10 references

Teaching signal synchronization in deep neural networks with prospective neurons

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Slow-integrating neurons equipped with an adaptive current can track an instantaneous target trajectory exactly, eliminating the cumulative delays that make teaching signals arrive out of sync in hierarchical networks.

desk verdict The math is the contribution: prospective dynamics provably eliminate tracking lag, but the biological implementation rests on an unverified fast-adaptation assumption. read the letter →

arxiv 2511.14917 v2 pith:LA5TWHUU submitted 2025-11-18 q-bio.NC cs.NE

Nicolas Zucchet , Qianqian Feng , Axel Laborieux , Friedemann Zenke , Walter Senn , João Sacramento This is my paper · ORCID
classification q-bio.NCcs.NE MSC 92B2037N25
keywords prospectiveneuronsteachingsignalsynchronizationadaptivecurrentsonlinelearningcreditassignmentleakyintegratornetworksequilibriumpropagationworkingmemory
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper claims that the temporal lag of slowly integrating neurons—which grows with network depth and makes error signals arrive out of sync with the activity they should modify—can be removed by giving each neuron a 'prospective' input that estimates what its input will be one time constant in the future. The mathematical core is a theorem showing that if the state follows prospective dynamics, the residual error between state and instantaneous input decays exponentially and the state eventually coincides with the ideal target trajectory, so the delayed network behaves like its instantaneous counterpart. The paper then shows that a biophysically plausible adaptive current, which subtracts a low-pass filtered copy of the input, approximates this prospective term and improves tracking whenever its time constant is fast enough. On that basis, the paper demonstrates that prospective dynamics synchronize teaching signals across backpropagation, feedback alignment, and equilibrium-propagation style rules, and that they enable online learning in a motor-control task and in a delayed-reach memory task. A sympathetic reader would care because this offers a single mechanism—neural adaptation—that could solve a critical timing problem for learning in the brain.

What carries the argument

The key object is the prospective dynamics equation: tau s_dot_t = -s_t + f_theta(s_t,t) + tau d_t f_theta(s_t,t), where the added term is a first-order prediction of where the input is heading. Rewriting it as tau d_t[s_t - f_theta(s_t,t)] = -(s_t - f_theta(s_t,t)) exposes the mechanism: the tracking error itself obeys an exponentially stable ODE, which is why the state locks onto the instantaneous target. The biophysical implementation is the adaptive current: a low-pass filtered copy of the input acts as a high-pass filter, and when the adaptation time constant tau_a is much smaller than the membrane time constant tau it reproduces the derivative term, up to corrections of order tau tau_a

What would settle it

In a single leaky neuron receiving a time-varying input, measure the steady-state lag between input and voltage. Leaky dynamics predict a constant phase lag growing like gamma tau / mu; prospective dynamics predict the residual error ||s_t - f_theta(s_t,t)|| decaying exponentially to zero and the voltage leading or matching the instantaneous target. If an adaptive neuron with tau_a < tau still shows a nonzero steady-state phase lag, or if the implicit Euler scheme in Eq. 17 diverges on a simple linear target where the ODE is stable, the central practical claim fails.

Watch

Extended reading notes

Core claim

The central discovery is that the delay problem for hierarchical learning is formally a tracking problem: neural activity should track the trajectory s*_t defined by s*_t = f_theta(s*_t,t), the instantaneous equilibrium of the same network. Leaky integrator dynamics tau s_dot = -s + f_theta(s,t) can approach s* only up to a tube of radius proportional to gamma tau / mu, so tracking error—and thus teaching-signal misalignment—grows with neuron time constant and input speed. Adding the prospective term tau d_t f_theta(s_t,t) changes the dynamics to tau s_dot = -s + f_theta(s,t) + tau d_t f_theta(s_t,t), and Theorem 2 shows that s_t - f_theta(s_t,t) decays as exp(-t/tau) and that limsup ||s_t -

Load-bearing premise

The biological implementation rests on the premise that an adaptive current can act as an accurate local differentiator, i.e., that the adaptation time constant tau_a is much smaller than the membrane time constant tau—a regime the paper itself notes lies outside where adaptive neurons traditionally operate; the numerical simulation scheme in Eq. 17 also has no theoretical convergence guarantees.

Editorial extensions

If this is right

  • Any learning rule whose updates require equilibrium activities and error signals—backpropagation, feedback alignment, direct feedback alignment, holomorphic equilibrium propagation, predictive coding—can in principle be run online in slowly integrating neurons with prospective dynamics, matching instantaneous-network performance.
  • Teaching signals no longer need to be delayed-matched with eligibility traces or forward models; synchronization is achieved at the level of neural dynamics themselves.
  • Memory can be kept in slow, non-prospective neurons while only their errors are prospective; the paper shows such a hybrid network learns a delayed-reach task that pure prospective or pure leaky networks cannot solve.
  • Reward-based online learning in near-continuous time (actor-critic on the inverted pendulum) becomes feasible with tau = 100 ms neurons, with graceful degradation under +/-10% time-constant mismatch and failure only as the prospective term vanishes.
  • The decoupling of activity and error dynamics suggests architectures with prospective errors but non-prospective activity, widening the class of biologically plausible learning rules.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the fast-adaptation premise can be met in real neurons (e.g., by fast sodium inactivation), the theory predicts a measurable phase advance of voltage responses relative to slowly varying inputs, and a corresponding improvement in temporally precise learning; this is a direct experimental handle.
  • The tracking theorem is not limited to feedforward networks; it applies to any equilibrium-defining dynamics with invertible Jacobian, so recurrent circuits and deep equilibrium networks should show the same synchronization benefit—an extension the paper notes but does not pursue experimentally for large recurrent models.
  • Complementarity with eligibility traces suggests a two-mechanism picture: traces delay the presynaptic term, prospective errors advance the postsynaptic term; combining them might extend online learning to substantially deeper recurrent networks than either mechanism alone.
  • A practical engineering translation is that a single first-order high-pass filter inside each processing unit could let online gradient-based learning replace batched or offline training in slow physical or analog hardware, provided the integration step is numerically stable.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper formalizes the temporal delay of neural activities relative to the instantaneous equilibrium trajectory as a tracking problem. It proves that leaky integrator dynamics have an asymptotic tracking error bounded by γτ/μ (Theorem 1) and that adding a prospective term τ d_t f_θ(s_t,t) to the dynamics makes the residual s_t - f_θ(s_t,t) decay exponentially (Theorem 2), so that prospective neurons can effectively synchronize teaching signals. The authors then propose that adaptive currents, implemented as a high-pass filter of the input current, provide a biologically plausible approximation of prospective dynamics, and they analyze time-constant mismatches. The framework is tested on a teacher-student task with several learning rules, an inverted-pendulum control task, and a delayed-reaching task with memory-storing leaky neurons trained by an online RTRL-type rule.

Significance. If the results hold, the paper makes a useful contribution: it gives a clean, parameter-free characterization of the tracking error of leaky neurons and of the exponential synchronization of prospective dynamics, generalizes earlier prospective-coding theories (neuronal least action, latent equilibrium) to non-energy-based systems, and demonstrates the benefit of prospective dynamics across several learning algorithms. The derivations in the appendices are mostly clear, the tightness construction for Theorem 1 is a nice addition, and the experiments span a broader set of tasks than most papers in this area. The strength of the paper is the theoretical core: Theorem 1 and the exponential identity of Theorem 2 are stated without hidden parameters and are validated by simulation. The biological and experimental claims are, however, less robust, as detailed below.

major comments (4)
  1. [Theorem 2 / Appendix A.3] The second part of Theorem 2, limsup ||s_t - s*_t|| = 0, does not follow from the stated assumptions. The exponential decay only shows that s_t - f_θ(s_t,t) → 0, i.e., that s_t approaches the set of fixed points of s = f_θ(s,t). Invertibility of ∂_s f_θ and Lipschitz continuity in t do not imply uniqueness of the fixed point or continuous selection of a branch; for example, f(s)=s^2 has two fixed points and the same finite-dimensional residual can decay while the state tends to either root. The intended feedforward/contractive setting is fine, but the theorem as stated is too broad. Please add the missing contraction/uniqueness assumption (e.g., the spectral condition used in Theorem 1) or prove convergence to the specific s*_t used in the paper.
  2. [Section 2.1 and Discussion 'Neural implementation of adaptive dynamics'] The biological implementation claim depends critically on the adaptation time constant τ_a being much smaller than the membrane time constant τ. The high-pass filter (f-a)/τ_a approximates the derivative only in the limit τ_a→0, and Eq. (8) makes the residual O(τ τ_a d_t^2 f) explicit. The paper itself states that τ_a < τ is 'outside the regime in which adaptive neurons traditionally lie' and that the hypothesis is 'critical to the theory of Section 2 and its experimental counterpart.' The only supporting evidence offered is a preprint (Brandt et al. 2024) with no direct measurement that effective τ_a is below τ. This does not invalidate the mathematical tracking theorem, but it means the advertised 'adaptive neurons can implement prospective dynamics' headline is conditional on an unverified biological assumption. Please either provide direct evidence or reframe the contribution as a th
  3. [Methods, 'Simulation of prospective dynamics'] The Euler-like implicit scheme in Eq. (17) is the backbone of all learning experiments, yet the authors state they are 'not aware of any theoretical guaranties' for it. Since prospective dynamics are described by a differential-algebraic equation, a naive scheme can be inaccurate or unstable in parameter regimes used in the paper, especially for the oscillatory hEP experiments. The comparison to Euler integration of the prediction-correction dynamics is reassuring but not reported quantitatively. Please add a convergence check against a high-accuracy solver for at least the teacher-student and cartpole setups, or provide an analysis of the scheme's consistency/stability.
  4. [Table 1 / Section 3.1] The experimental evidence for the 'large variety of learning rules' claim rests on Table 1, which reports point estimates with no error bars. The learning rate is grid-searched separately for each method on only 2 seeds, so the ranking of methods may partly reflect per-method hyperparameter overfitting. For the recurrent/hEP results, the oscillation parameters |β|, ω_β, and τ_ema are also chosen by grid search. Please report mean and variance over at least 5–10 seeds, and state whether the qualitative ranking (prospective > leaky, BP ≈ prospective) is stable across the grid. This is not fatal to the theory, but it is load-bearing for the claim that prospective dynamics transfer across learning rules.
minor comments (5)
  1. [Methods, Section 3.1] The text says 'For the recurrent BP and holomorphic EP learning experiments (Figure 1)', but the recurrent results are shown in Table 1, not Figure 1. Please correct the reference.
  2. [Theorem 2 statement] The assumption that f_θ is 'Lipschitz continuous in t' should be stated more precisely, e.g., with respect to the norm on the state space and uniformly for the relevant trajectory; this will also clarify the proof of the limsup claim.
  3. [Figure 3] The panels E and F show time-constant mismatch; please add a legend or caption detail clarifying whether τ' is varied on the x-axis and in which units (multiples of τ). The current text refers to 'the negative peak around 0' in a way that is hard to verify from the figure.
  4. [Typography] Table 1 contains 'T rain loss' and 'T est loss'; the abstract contains 'Correspondance'. These typos should be fixed.
  5. [Section 2.2] The statement 'the prospective and leaky dynamics match for τ'=0' is correct but slightly misleading, since the mismatch analysis is local around τ'=τ. It would be helpful to note explicitly that the comparison to leaky dynamics at τ'=0 is outside the range where Eq. (11) is guaranteed to hold.

Circularity Check

0 steps flagged · score 2.0 of 10

No substantive circularity: the central tracking theorem is a self-contained derivation; the biological implementation rests on an openly flagged, partly self-cited fast-adaptation assumption, which is a limitation rather than a circular step.

full rationale

The mathematical chain is self-contained. Theorems 1 and 2 are proved from the stated ODEs (Eqs. 3 and 5) with no fitted parameters; Theorem 2 is a direct algebraic consequence of the defining prospective dynamics, i.e. the mechanism is explicitly constructed so that the residual s_t - f_theta(s_t,t) decays. That is a design/construction, not a hidden equivalence between a fitted input and a predicted output. The paper also openly attributes prospective dynamics to prior work (Senn et al. 2024; Haider et al. 2021) and shows in Methods that Eq. 5 is the same as the known prediction-correction algorithm (Eq. 15), so the connection is disclosed rather than renamed as new. The main circularity-adjacent weakness is biological: Eq. 7 yields prospective behavior only when tau_a << tau, and the Discussion states this is 'critical to the theory of Section 2 and its experimental counterpart' and 'outside the regime in which adaptive neurons traditionally lie'; the only mechanistic support invoked is a same-group preprint [Brandt et al. 2024] on sodium-inactivation fast adaptation. This is load-bearing for the neural-implementation claim but is an openly stated empirical assumption, not a derivation that reduces to its inputs; the abstract tracking theorem remains valid without it. Likewise, the simulation scheme (Eq. 17) is admitted to lack theoretical guarantees; this is a limitation, not circularity. Self-citations are used for context or components and are not the sole support of any derivation. Accordingly, no prediction reduces to a fit or to a self-citation chain; circularity is minor.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The theoretical core relies only on standard smoothness/contraction assumptions plus the domain assumption that the instantaneous equilibrium is the desired target. The biological claim carries an additional domain assumption about fast adaptive currents, and the simulation results depend on hyperparameters and a numerically unproven integration scheme. No exotic new entities are introduced.

free parameters (3)
  • Per-method learning rate = grid search over [0.001,0.003,0.01,0.03,0.1,0.3,1]
    In Section 3.1 Methods, the learning rate is tuned independently per method using 2 seeds; the best average test loss is used for plots. This affects reported learning performance but not the theoretical tracking theorems.
  • hEP oscillation parameters |β|, ω_β, τ_ema = |β|=0.2, ω_β=190π rad/s, τ_ema=31.5 ms
    Grid-searched for the holomorphic equilibrium propagation experiment; the Table 1 hEP results depend on these choices.
  • Cartpole reward-region margins = -2≤x≤2, -8.5°≤θ≤8.5°
    Narrower than the termination range; the paper states this change was 'critical' for the value function to learn the episode end. This is post hoc reward shaping.
assumptions (5)
  • domain assumption Contraction: largest eigenvalue of the symmetric part of ∂_s f is ≤ 1-μ for μ>0 (Theorem 1)
    Used to prove the leaky-tracking bound; a property assumed of the network dynamics, not derived.
  • domain assumption Invertibility of ∂_s f and Lipschitz continuity in t along the trajectory (Theorem 2)
    Needed for the prospective dynamics to define a trajectory and for the target s* to be well-defined and continuous.
  • domain assumption Uniqueness/existence of the implicit target trajectory s*_t satisfying s* = f(s*,t)
    The paper notes multiple target trajectories can exist for recurrent f, but Theorem 2's convergence claim requires the target to be well-defined.
  • domain assumption High-pass filtered input approximates d_t f when τ_a→0; adaptation can be fast in biological neurons
    Central to the adaptive-neuron implementation; the paper concedes τ_a<τ is outside the standard adaptive-neuron regime and invokes speculative sodium-inactivation mechanisms.
  • ad hoc to paper Stability of the ad-hoc Euler-like integration scheme Eq. 17
    The authors state they are not aware of theoretical guarantees for this scheme; simulations rely on it empirically.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Teaching signal synchronization in deep neural networks with prospective neurons." pith.science (2026). https://pith.science/paper/LA5TWHUU

@misc{pith2026251114917,
  author       = {Pith},
  title        = {Pith review of: Teaching signal synchronization in deep neural networks with prospective neurons},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LA5TWHUU}},
  note         = {Machine review of arXiv:2511.14917}
}
read the original abstract

Working memory requires the brain to maintain information from the recent past to guide ongoing behavior. Neurons can contribute to this capacity by slowly integrating their inputs over time, creating persistent activity that outlasts the original stimulus. However, when these slowly integrating neurons are organized hierarchically, they introduce cumulative delays that create a fundamental challenge for learning: teaching signals that indicate whether behavior was correct or incorrect arrive out-of-sync with the neural activity they are meant to instruct. Here, we demonstrate that neurons enhanced with an adaptive current can compensate for these delays by responding to external stimuli prospectively -- effectively predicting future inputs to synchronize with them. First, we show that such prospective neurons enable teaching signal synchronization across a range of learning algorithms that propagate error signals through hierarchical networks. Second, we demonstrate that this successfully guides learning in slowly integrating neurons, enabling the formation and retrieval of memories over extended timescales. We support our findings with a mathematical analysis of the prospective coding mechanism and learning experiments on motor control tasks. Together, our results reveal how neural adaptation could solve a critical timing problem and enable efficient learning in dynamic environments.

Figures

Figures reproduced from arXiv: 2511.14917 by the authors.

Figure 1
Figure 1. Prospective dynamics can track the target trajectory whereas leaky dynamics cannot. Under the leaky integrator dynamics, the current state st is attracted towards the current target s ∗ t . By the time it arrives there, the target will have already moved away, so s is always lagging behind s ∗ . Adding a prospective input to the leaky dynamics enables perfect tracking, as the position of s ∗ in the near future is in… view at source ↗
Figure 2
Figure 2. Theoretical predictions on tracking properties of leaky and prospec￾tive neurons hold in simulation. A. Empirical tracking error for the leaky dynamics (3) as a function of the neurons time constant τ times the characteristic angular ve￾locity of the inputs ω0. This quantity is directly linked to the theoretical bound of Theorem 1. Each dot corresponds to random τ and ω0 values. As predicted by The￾orem 1, the track… view at source ↗
Figure 3
Figure 3. Better target trajectory tracking translates to greater learning performance. Top row measures the empirical tracking error in the setup of Sections 1 and 2, bottom row the test loss in the teacher-student task of Section 3.1. Note that the two setups are similar but have some differences, cf. Methods. A, B. Leaky dynamics (3) require τ → 0 to perfectly track the reference trajectory and thus effectively learn (blue… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Prospective dynamics support learning in a control task. A. Visual depiction of the inverted-pendulum task, in which the goal is to balance the pole in an upright position. Reward rate is 1 as long as the pole remains in the “target angles” region and that cart is not …
Figure 5
Figure 5. Figure 5: Prospective neurons support learning of memory-storing non￾prospective leaky neurons. A. We consider a multi-layer neural network combining complex-valued leaky neurons that integrate information over time (first layer) with prospective neurons (in the subsequent two l…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 1 linked inside Pith

  1. [1]

    Finally, we updateθonline, using the neural activityu t and error signalδ t currently available

    Instead, we apply the prospective dynamics to always track equilibrium. Finally, we updateθonline, using the neural activityu t and error signalδ t currently available. This yields ourProspective BPalgorithm:    τ˙ut =−u t +W t ρ(ut) +τ ˙ρ(ut) τ ˙δt =−δ t +ρ ′(ut)W ⊤ t δt +τ d dt ρ′(ut)W ⊤ t δt τW ˙Wt =−δ tρ(ut)⊤ (34) In our simulations of Sectio...

    1986

  2. [4]

    exp(−t/2) + 2t−4 so∥s ∗ t −s t∥ →4. Remarks.Let us remark on several things: •The result and the proof are inspired by a result from [Polyak, 1987, Chapter 6] for time-varying optimization, which assumesf(s, t) =∇sE(s, t). We extend this result to generalf, which requires changing theµstrong convexity assumption needed in the time-varying optimization set...

    1987

  3. [8]

    •In the main text, we claimed that ˙s ∗ t both increases with how fast external input varies and depends on the geometry of∂ sfθ

    A discrete-time version of this result can be obtained by adapting the proof of Popkov [2005]. •In the main text, we claimed that ˙s ∗ t both increases with how fast external input varies and depends on the geometry of∂ sfθ. This is because the implicit function 33 theorem applied tos ∗ t −f θ(s∗ t , t) = 0 gives ˙s∗ t =− Id− ∂fθ ∂s (s∗ t , t) −1 ∂fθ ∂t (...

    2005

  4. [9]

    The second one leverages the Taylor expansion of dtfθ(ut, t). For it to be mathematically rigorous, one must assume that the third-order time derivative off θ(ut, t) is uniformly bounded and make use of the Lagrange formulation of Taylor’s remainder. Finally, the last one comes from the standard integral values 1 τa Z ∞ 0 (t′ −t) exp − t−t ′ τa dt′ =τ a (...

    2022

  5. [10]

    have shown that oscillations can exactly compute the gradient in the weak nudging limit. Finally, we argue that the separation of timescales needed here is less restrictive than the one assuming infinitely fast neurons; here, only the teaching signal, a small part of the network, has to be fast. D Additional simulation details for the inverted pen- dulum ...

    2022

  6. [1952]

    Prospective and retrospective coding in cortical neurons.arXiv preprint arXiv:2405.14810,

    Simon Brandt, Mihai Alexandru Petrovici, Walter Senn, Katharina Anna Wilmes, and Federico Benitez. Prospective and retrospective coding in cortical neurons.arXiv preprint arXiv:2405.14810,

    arXiv

  7. [2002]

    Petrovici

    27 Benjamin Ellenberger, Paul Haider, Jakob Jordan, Kevin Max, Ismael Jaras, Laura Kriener, Federico Benitez, and Mihai A. Petrovici. Backpropagation through space, time, and the brain.arXiv arXiv:2403.16933,

    arXiv

  8. [2013]

    A theoretical framework for back-propagation

    Yann LeCun. A theoretical framework for back-propagation. InProceedings of the 1998 Connectionist Models Summer School,

    1998

Show all 10 references
  1. [2022]

    URLhttp://github.com/RobertTLange/gymnax. 29 A Theoretical derivations A.1 Derivation of the error backpropagation dynamics in con- tinuous time In Section 1.1, we claimed that the dynamics (3) compute the error signal of backprop- agation. We prove this statement by leveragin...

    1988

  2. [2024]

    Lapicque’s introduction of the integrate-and-fire model neuron (1907)

    Larry F Abbott. Lapicque’s introduction of the integrate-and-fire model neuron (1907). Brain Research Bulletin, 50(5-6),

    1907

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.