acceptodds
Under review as a conference paper at ICLR 2027

Does Feedback Alignment Work at Biological Timescales?

Abstract

Feedback alignment and related weight-transport–free algorithms of error propagation are often proposed as biologically plausible alternatives to backpropagation, yet they are typically formulated in discrete phases with implicitly synchronized forward and error signals. We develop a continuous-time model of feedback-alignment–type learning in which neural activities and synaptic weights evolve together under coupled first-order dynamics with distinct propagation, plasticity, and decay time constants. Our analysis shows how temporal overlap between presynaptic drive and locally projected error signals governs robustness to timing mismatch. For these two-signal rules, slow weight dynamics cannot bridge feedback delays after stimulus-specific activity has disappeared. We demonstrate learning with millisecond neural dynamics and stimulus presentations lasting tens to hundreds of milliseconds, with simultaneous inference and plasticity.

Primary area:
applications to neuroscience & cognitive science
Keywords:
feedback alignment, biologically plausible learning, continuous-time learning, credit assignment

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yzEfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a model of feedback alignment (FA) in continuous time. It is a coupled ODE framework where neural activities and synaptic weights evolve simultaneously, without separating inference and learning into different phases. The authors derive a temporal-overlap law to describe how delays between input activity and feedback error signals affect synaptic updates. Experiments on downsampled MNIST and synthetic classification tasks show that the proposed method can learn with millisecond neural dynamics, while longer delays and slower neural propagation generally reduce learning performance.

Critical strengths

The motivation for studying FA in continuous time is understandable. The proposed framework provides a way to study these timescales within the same dynamical system.

Critical weaknesses
  1. Although the motivation for introducing a continuous-time formulation is reasonable, I am not fully convinced about its novelty and necessity. What is the main difference between the proposed continuous-time FA and conventional discrete-time FA with explicitly modeled delays? What new insights or properties can be obtained from the continuous-time formulation that cannot be studied using discrete-time models? These questions are not clearly presented in the paper.
  2. The temporal-overlap law is presented as a main theoretical contribution. Does this law provide any specific insights into the learning dynamics or performance of FA? The experimental results qualitatively agree with the overlap principle, but I do not see a clear quantitative connection between the theoretical results and the actual learning performance.
  3. The authors compare their method with discrete-time SGD, but a direct comparison with discrete-time FA under similar delay conditions would be more relevant.
  4. The paper emphasizes biological plausibility, but I think it is important to distinguish temporal realism from biological plausibility more carefully. Introducing continuous neural dynamics and biologically relevant time constants makes the model more realistic in terms of timing, but this does not establish that the learning mechanism is biologically plausible. I think the authors should be more careful in framing and supporting their claims about biological plausibility.
Questions for authors

The proposed method learns the feedback weights through an additional plasticity rule, this is different from discrete FA. Is it necessary to learn the feedback weights? Does this improve learning performance or robustness to delays?

4 comments

Thank you for the review. We do not propose a better model: we investigate an existing one, feedback alignment, in a regime where inference and plasticity run at the same time and the teaching signal can be late.

CW1. "What is the main difference between the proposed continuous-time FA and conventional discrete-time FA with explicitly modeled delays? What new insights or properties can be obtained from the continuous-time formulation that cannot be studied using discrete-time models?"

A discrete-time model does not capture the finite speed at which information travels: each update pairs an input with its own error at the same instant. The continuous-time formulation includes it, and it shows three things a discrete model cannot:

  1. Late teaching signals. How much of each update pairs an input with its own label follows from the timing itself: a fraction g(Δ) = (1 − |Δ|/T)₊ (Eqs. 3–5), which we derived and now test quantitatively (CW2). A discrete model would have to assume it.
  2. Which time constants can compensate. Slower or faster plasticity cannot make up for a delay, because plasticity and decay act only after the input and error have met (Eq. 2). A discrete model, with a learning rate but no time course, cannot ask this.
  3. Finite propagation speed. Propagation through the layers adds to the delay, and learning degrades as the propagation constant approaches the presentation time (Fig. 7).

CW2. "Does this law provide any specific insights into the learning dynamics or performance of FA? … I do not see a clear quantitative connection between the theoretical results and the actual learning performance."

The law says when FA can learn under timing mismatch, and it predicts how well (new Appendix E.1, Fig. 9):

  1. Collapse. In Figs. 5 and 6, accuracy depends on delay and presentation length almost only through |Δ|/T (T = hold + ramp): this one ratio carries 92–98% of the variance that delay and length carry together, on both tasks and at all depths.
  2. Prediction with nothing fitted. A discrete version of the same rule (batch size one; the same seeds, data order and initial weights; step size T/τ_plas; each update weighting the correct label by g and the neighbour's by 1 − g) reproduces the continuous-time accuracy for presentations of 40 ms and longer: the rms difference over (delay, length) cells is 0.004–0.014.
  3. Where learning fails. If a fraction 1 − g of updates carries a label drawn from the class prior, a learner keeps a minority class only while g exceeds a threshold set by the class balance. This is why the 75/25 circles task falls half-way to chance at |Δ|/T = 0.33–0.49 while ten-class MNIST holds to about 0.9, and why changing only the class prior in the discrete model moves the threshold. Section 5.2 now explains Fig. 6 this way.

CW3. "a direct comparison with discrete-time FA under similar delay conditions would be more relevant."

We agree, and the revised Fig. 8 (Appendix D) makes this comparison, matched sample by sample: the circles networks trained discretely with the same seeds, data order and initial weights, each update mixing the correct and the neighbouring label in the proportion g(Δ) that our analysis derives from the timing. With this mixing, discrete back-propagation loses accuracy with delay just as our continuous-time networks do (it sits slightly above them, by 0.015–0.034), and so does the same rule run discretely (within 0.012 on average). This is what the comparison shows: delay tolerance is set by the pairing fraction g(Δ), not by the learning rule, and a discrete model reproduces it only once it is given g, which comes from the continuous-time analysis. The continuous-time comparison with fixed random feedback is in our answer to Q1.

CW4. "it is important to distinguish temporal realism from biological plausibility more carefully."

We agree and adopt your distinction: the paper tests the temporal realism of an existing rule, not whether FA is how the brain learns. The revised Introduction says so (Section 1).

Q1. "Is it necessary to learn the feedback weights? Does this improve learning performance or robustness to delays?"

Learning the feedback weights is not necessary for learning, and our timing results hold either way. We ran the paper's networks with the feedback weights held fixed at a random start, as in standard FA and DFA (Appendix E.2, Table 3). Both versions learn similarly well (at zero delay 0.93–0.98 on the circles with fixed feedback against 0.91–0.97 with learned feedback, and 0.88 against 0.86 on MNIST) and fall half-way to chance at similar delays; learned feedback started from a random matrix does about as well as or better than fixed feedback on the circles. We study the learned form because it includes the Kolen–Pollack and weight-mirror variants, whose learned feedback aligns with the forward weights and comes close to back-propagation at scale in discrete time (Akrout et al. 2019).

I thank the authors for these answers.

I have a further question about what the authors think about the real role of feedback weights: right now the learned and random feedback weights exhibit same performance, which is non-intuitive to understand.

Thank you for the follow-up. For the question our paper asks, the choice does not matter, and our analysis predicts this: a delay corrupts which label is paired with each input before the feedback weights come into play, so delay tolerance is the same for learned and fixed feedback. Their equal accuracy is also expected. With learned feedback the feedback weights align with the forward weights; with fixed random feedback the forward weights align with the feedback instead, the original feedback-alignment effect. On 7×7 MNIST, the alignment rises from about 0 to 0.84 with learned feedback and to 0.71 with fixed feedback, so the hidden layers receive a useful error either way. Learned feedback also aligns layers further from the output, which matters more for deep networks and large tasks (Akrout et al. 2019) than for ours.

Reviewer_3hg1OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2(+1)
Read review
Paper summary

The paper builds on the feedback alignment literature and attempts to establish properties requisite for continuous-time versions of feedback alignment. The paper starts by defining a set of differential equations governing the neural states, feedforward weights, and feedback weights. Using these equations, which each have their own time constant, the authors establish that for synaptic learning to occur there must be a temporal overlap of the feedback signal, and further the authors establish that synaptic learning can happen in biological timescales. The authors validate their claims by performing experiments on a 7x7 downsampled MNIST dataset on a 1-hidden-layer neural network.

Critical strengths

None

Critical weaknesses

The critical weaknesses of this paper can be enumerated under the common denominator of the strength and originality of the contribution. Foremost, the paper seems to construct a question that has already been solved and much expanded upon, in combination with a feeble empirical support for these reiterated questions. Continuous-time feedback-alignment-like models have been around for a decade and the paper’s introduction and related work do not consider models created after 2019, nor consider any literature that concern themselves with the biological timescales and timeconstants that the paper works on. For example, the concept of a continuous-time feedback alignment model has readily been introduced (Meulemans 2021,2022) in the Deep Feedback Control framework, insofar that Eq. (1) from the paper is a literal copy of the equations from this paper. Continuous-time feedback weight learning has also been broadly explored in the further extensions of target propagation (the follow-ups of Bengio’s work), predictive coding (Bogazs, Buckley), and other biologically plausible circuitry (Ponte Costa). Further, the kernel of Eq. (5) is a rate-based estimate of the previously-established learning rules derived from actual pyramidal neurons (Aceituno 2024, 2023). More broadly, analysis of timescales and time constants has been done before (e.g. Moro 2026, etc), and the paper does not seem to actually specify what timescales might be important for brain research (Zeraati 2026). Incidentally, the claim " learning in biology unfolds continuously in time, and does not rely on separate learning and inference phases" is hard to sustain given the wealth of evidence showing that offline cortical reactivations and hippocampal replays are necessary for learning (e.g. Nghia 2024, Lensjo 2026). Another critical weakness is the empirical support of the work. Not only are the networks trained extremely small (49/32/24 neurons with one hidden layer) also the dataset is highly nonstandard as a 7x7 downsampled MNIST would not normally be considered as a task or a problem, as you can solve normal MNIST with small reservoir computing, and the current state of the field is training ImageNet with continuous-time networks (Kerjan 2026). All in all the idea is simply not a large or original contribution.

Questions for authors
  • Why did you not consider any literature on timescales in dynamical ANNs?
  • What would you explain as the major contribution in this work that is not present in other works?
  • Why only perform experiments on absurdly small problems?
4 comments

Thank you for the detailed pointers; we agree that the related work needs the additions below. But the central claims of the review rest on misreadings of the cited work, and we would like to first correct the most serious one.

CW2. "the concept of a continuous-time feedback alignment model has readily been introduced (Meulemans 2021,2022) in the Deep Feedback Control framework, insofar that Eq. (1) from the paper is a literal copy of the equations from this paper"

The review's claim that we copied our equations from Deep Feedback Control (DFC) is false: DFC is a completely different algorithm, and the only thing the two models share is the textbook leaky-integrator neuron. That neuron is a standard leaky-integrator rate neuron (Eq. 7.9 of Dayan & Abbott (2001)); by the reviewer's standard, every leaky-integrator rate network would be a copy of DFC. In DFC a PI controller drives every neuron's activity toward a target; the forward weights learn by comparing a neuron's controlled activity with its feedforward input; and the feedback weights learn from injected noise (Meulemans et al. 2021, 2022). None of this exists in our model: there is no controller, feedback never changes a neuron's activity but only gates plasticity, and both weight sets learn from products of forward activity and error. Nor is DFC feedback alignment: it approximates Gauss–Newton updates, which its authors call "fundamentally different from BP approximations", and it uses DFA as a baseline. The revision cites both papers and states this difference (Section 2).

CW1, Q2. "the paper seems to construct a question that has already been solved and much expanded upon" … "Continuous-time feedback-alignment-like models have been around for a decade"

The model the review gives as its example, DFC, is not feedback alignment (above). Other learning algorithms with continuously evolving activity exist, such as dendritic microcircuits (Sacramento et al. 2018) and Latent Equilibrium (Haider et al. 2021), but they are not feedback alignment either. FA itself was run by the original FA paper with spiking units whose forward and feedback pathways are active at once (Lillicrap et al. 2016, Fig. 4). The revision discusses all of these (Sections 1 and 2). None of the existing literature asks our questions: when inference and plasticity run at the same time, how much timing error between an input and its teaching signal can FA tolerate, what sets that limit, and can a different plasticity rate make up for it? None of them varies the offset between an input and its teaching signal; in Lillicrap et al. the target switches with the image, which is the zero-delay case of our experiments.

CW3. "Continuous-time feedback weight learning has also been broadly explored in the further extensions of target propagation …, predictive coding …, and other biologically plausible circuitry"

These are different learning algorithms, and our paper is about feedback alignment; we make no claim about target propagation, predictive coding or DFC. That some of them learn feedback weights in continuous time does not bear on our question, because none of them studies how an offset in time between an input and its teaching signal affects learning. That is the question we answer, for FA. Section 2 now states this distinction.

CW4. "the kernel of Eq. (5) is a rate-based estimate of the previously-established learning rules derived from actual pyramidal neurons (Aceituno 2024, 2023)"

Neither paper contains it. The kernel of Eqs. 2–4 is the solution of the weight-decay term and acts after the coincidence product is formed; g(Δ) in Eq. 5 is the overlap of two boxcars of the stimulus duration, not a synaptic window. Aceituno et al. 2023 use Δw ∝ ∫ r_pre ṙ_post dt, Aceituno et al. 2024 a calcium- and burst-based three-factor rule. Biological plasticity rules do filter a factor before the product (STDP windows, eligibility traces). FA does not, and we study FA as it is: this is a difference between FA and biology, not a limitation of our analysis.

CW5, Q1. "analysis of timescales and time constants has been done before (e.g. Moro 2026, etc), and the paper does not seem to actually specify what timescales might be important for brain research (Zeraati 2026)"

Moro et al. (2026) and Zeraati et al. study neuronal and activity timescales, not learning-rule or feedback timing. Our analysis identifies which timescales decide whether FA can learn: the delay between an input and its teaching signal relative to how long the input is held, |Δ|/T (Section 4). The plasticity and decay time constants do not change how the update is paired, so they cannot make up for a delay, and propagation adds to the delay; with 50 ms presentations we show learning for propagation constants up to about 5 ms (Section 5.3). Where in the brain FA might run, and so which of its measured timescales apply, is not something the paper claims; any such mapping would have to check this ratio. The revision cites them, with the closest analyses of timing in continuous-time learning (Latent Equilibrium; Senn et al. 2024; Fayyazi et al. 2024; Ralambomihanta et al. 2026), in Section 2.

CW6. "offline cortical reactivations and hippocampal replays are necessary for learning (e.g. Nghia 2024, Lensjo 2026)"

Our results concern learning that happens while the animal behaves, with inference and plasticity running at the same time. Such online learning is well documented: hippocampal place fields form in a single trial (Bittner et al. 2017), and cerebellar motor learning pairs ongoing activity with error signals that arrive about 100 ms later (Raymond & Lisberger 1998). Offline replay may add consolidation on top of this, but the cited papers do not show that it is necessary for learning in general: in Lensjø et al. (Science Advances, 2025) performance improves within training sessions, and what their manipulation impairs is consolidation across days. The revised Introduction says that we model this online component (Section 1).

CW7, Q3. "the networks trained extremely small … the current state of the field is training ImageNet with continuous-time networks (Kerjan 2026)"

Kerjan et al. 2026 use equilibrium propagation to train a model on ImageNet. This is not only a different algorithm than FA, but it also uses two nudged phases, ten discrete state iterations per phase, a separate parameter-update phase, feedback through the transpose of the forward Jacobian, and momentum SGD on a GPU: that is not a continuous-time, phase-free or weight-transport-free network. Its learning depends on precisely timed phases, and it never faces the problem we study: an input and its teaching signal that do not arrive together. Our small tasks kept dense delay × duration grids affordable; the derivation of the law does not depend on them.

We hope you will reconsider the contribution for what it is: a quantitative account, tested against simulation, of how much timing error the FA-type rule tolerates, which to our knowledge no prior work provides.

Dear authors,

CW2. I accept the correction “literal copy” was inaccurate. The models share leaky-integrator neurons with coupled weight dynamics on separated timescales (the same setup underlies DFC, Latent Equilibrium and the dendritic microcircuit models). The contribution therefore has to come from the timing analyses.

CW1/Q2. The revised results show that the effect is not specific to feedback alignment as backpropagation with the same label-mixing degrades in the same way. The authors also show that ordinary discrete-time training reproduces the continuous-time accuracy with nothing fitted, when each update weights the correct label by g(Δ) = 1−|Δ|/T and the neighbouring sample’s label by the rest. A timing offset thus turns a fraction |Δ|/T of the updates into label noise, for any learner, and the failure threshold is the standard condition for learning under class-conditional label noise (Natarajan et al. 2013). Phase-free learning without weight transport is also already established (Neftci et al. 2017; Greedy et al. 2022; Max et al. 2024). A contribution specific to FA would show how offsets affect the alignment between forward and feedback weights, or where FA and backpropagation diverge wrt delay.

CW3. I agree that these are different algorithms, although the authors are aware of the shared literature and idea-space. When both W and V learn, which the paper itself describes as a Kolen–Pollack or weight-mirror-style scheme, not identical to FA or DFA. For continuous-time, phase-free learning of feedback weights, the closest prior work is Phaseless Alignment Learning (Max et al. 2024), in which all weights learn at the same time with always-on plasticity. Should be discussed and the claims should be stated for the learned-feedback rule.

CW4. I accept that the kernel does not appear exactly in the cited papers, I was referring to the rate-based learning rule in that paper that also uses the same overlap counting argument, but I wont press this point. However, “biological plasticity rules filter a factor before the product is formed (STDP windows, eligibility traces)“, and the paper’s rule does not. Thereby paper’s insight is that no choice of τ_plas or τ_dec can recover an association once stimulus activity has ended, which doesnt follow from biological timescales but directly from leaving out such a filter. If teaching signals in cortex arrive on behavioural timescales (see the reply on CW5), the paper’s own argumentation implies that the trace-free rule cannot learn there.

CW5. The point of both references is to explain that the timescales that are used to process information in a hierarchical network are given by the network, not by a broad notion of what is behaviourally observed. The delay between teaching signal and input (mentioned in the author response) is likely long and mediated by behavioural timescale plasticity (observed in hippocampus, again consistent with the complementary learning system framework). The timescales by which a layer in the cortical hierarchy learns is likely tied to the timescales of the next level in that hierarchy.

CW6. There is indeed evidence that the brain can do online learning, and in fact the complementary systems framework requires to some extent online learning. The response seems to miss the point that there is a difference between brains and neocortex (note that the papers I mentioned are the latest evidence of offline learning in the neocortex). The online learning that the authors refer to is hippocampal and cerebellar, both of which have not hierarchical representations (or "deep" as they would be called in machine learning). FA was introduced to explain that the formation of hierarchies in cortex could be done without weight transport. Thus, the e neuroscientific literature indicates that this learning happens offline, with compressed timescales (see prev. answer).

CW7/Q3. My concerns on the experiments remain. On 7×7 MNIST a linear classifier already reaches about the same accuracy as the networks, so the only task that actually requires hidden layers is a 2D toy problem.

I accept that my initial review was too harsh, but I think that some critical problems remain. I will update the soundness, but maintain my recommendation of a rejection.

The main issue is that the question posed by the paper does not seem, in my opinion, relevant for either machine learning nor computational neuroscience. Studying the FA from a machine learning perspective would require: more relevant benchmarks, and to explain when and why is FA an interesting algorithm (for ML), at those timescales. From the computational neuroscience perspective, the question requires a better justification. FA was interesting as a proof that the weight transport problem could be sidestepped in cortical hierarchical learning, but the examples of learning with the timescales that the authors study are non-cortical (more specifically, they don't require hiearchical representations), which then undermines the choice of FA.

Reviewer_vyJQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates whether feedback alignment can support continuous-time learning, in which neural activity and synaptic plasticity evolve simultaneously. The authors develop a coupled dynamical framework with separate timescales for neural propagation, plasticity, and weight decay. The central theoretical result is that effective learning depends on the temporal overlap between presynaptic activity and local error signals. The authors demonstrate the effectiveness of their method on downsampled MNIST and concentric-circle classification tasks.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. The paper identifies temporal overlap between neural activity and error signals as a key constraint on continuous-time learning, providing an intuitive explanation for the effects of propagation delays.
  3. The study provides a more biologically plausible framework for feedback alignment.
Critical weaknesses
  1. My main concern is that the study focuses on relatively simple network architectures and classification tasks, whereas biological neural circuits involve recurrent connections and must learn temporal relationships between events. It remains unclear whether the proposed framework can support even simple sequence-learning tasks, particularly when inputs and their associated errors are separated in time. I understand that the primary goal of this study is to demonstrate the feasibility of continuous-time feedback alignment. Nevertheless, extending the evaluation to simple recurrent networks and temporal prediction or sequence association tasks would provide stronger evidence for the generality of the proposed framework.

  2. As the authors acknowledge, biological synapses exhibit mechanisms such as eligibility traces that can preserve information about past neural activity. These mechanisms may enable delayed error signals to drive synaptic updates without direct temporal overlap. Although the authors leave this question for future work, even a preliminary investigation of such mechanisms would help establish whether temporal overlap remains a critical constraint in more biologically realistic learning systems.

  3. Beyond biological plausibility, what additional advantages does the proposed continuous-time framework offer over conventional discrete-time learning methods? Although the reported results demonstrate comparable performance and some robustness to feedback delays, it remains unclear whether the framework provides broader benefits in learning efficiency, convergence speed, final accuracy, or robustness. A more systematic comparison would help clarify its practical value beyond biological plausibility.

Questions for authors
  1. Beyond biological plausibility, what practical advantages does the proposed continuous-time framework offer? Compared with conventional discrete-time feedback alignment (FA), does it provide measurable improvements in learning performance or computational efficiency?

  2. If feedback weights are kept fixed rather than learned, how would this affect learning performance and stability under the same continuous-time dynamics?

  3. How robust is the proposed method to different initializations of the forward and feedback weights? In particular, how do these initializations affect learning performance, stability, and convergence?

3 comments

Thank you for a careful review. To clarify, we are not proposing a new or better learning rule: we investigate an existing, widely used one, feedback alignment (with its DFA and Kolen–Pollack variants), in a new and relevant regime: inference and plasticity running at the same time, with an input and its teaching signal that need not arrive together. To our knowledge, no prior work has asked how much timing error between the two this rule tolerates. We imposed controlled offsets, derived the fraction of each update that pairs an input with its own teaching signal, g(Δ) = (1 − |Δ|/T)₊ for an offset Δ and a presentation period T, and showed that this fraction predicts accuracy across delays and presentation lengths.

CW1. "the study focuses on relatively simple network architectures and classification tasks … It remains unclear whether the proposed framework can support even simple sequence-learning tasks, particularly when inputs and their associated errors are separated in time."

Our paper studies feedback alignment (FA) learning algorithms. Sequence learning (learning tasks with a temporal ordering) is a different task from the one FA and the model class we study were built for, so it is not the subject of this paper. Learning when input and error signals do not overlap in time, as such tasks require, needs something like an eligibility trace or persistent activity: our analysis shows that when presynaptic activity and the feedback error do not overlap there is no learning, no matter the plasticity and decay time constants (Eq. 2). That is exactly what other models of sequence learning add (RFLO, Murray 2019; e-prop, Bellec et al. 2020, now cited in Section 2), and it is outside the scope of our study.

CW2. "even a preliminary investigation of such mechanisms would help establish whether temporal overlap remains a critical constraint"

Our results suggest the need for persistent activity or a trace, but we do not argue that either alone is sufficient. A trace changes the learning rule, so testing it means studying a different model; our paper studies feedback alignment. The need for a trace under delay is the classical distal-reward problem (Izhikevich 2007; Gerstner et al. 2018), now cited in Section 2.

CW3, Q1. "what additional advantages does the proposed continuous-time framework offer over conventional discrete-time learning methods?" … "measurable improvements in learning performance or computational efficiency?"

We are not trying to "beat" other models. Our contribution is to test an existing model under timing mismatch, which had not been done, and to derive a law that predicts the outcome. As Appendix D says, the comparison with discrete training is a sanity check that the rule learns, "not to argue that our continuous-time networks outperform vanilla SGD". New in this response, we tested that law quantitatively (Appendix E.1; details in our reply to Reviewer yzEf): accuracy in Figs. 5 and 6 depends on delay and presentation length almost only through |Δ|/T, and a discrete version of the same rule with the law's label mixing reproduces the continuous-time accuracy with nothing fitted.

Q2. "If feedback weights are kept fixed rather than learned, how would this affect learning performance and stability …?"

We ran this (Appendix E.2, Table 3): the paper's networks, delays and presentation lengths, with the feedback weights held fixed at a random start. On the circles networks this is standard FA, on 7×7 MNIST standard DFA. Fixed feedback learns similarly well: at zero delay it reaches 0.93–0.98 on the circles against 0.91–0.97 with learned feedback, and 0.88 against 0.86 on MNIST, and it falls half-way to chance at the same or longer delays. Learned feedback started from a random matrix does about as well as or better than fixed feedback on the circles networks, and the overlap law describes every version. The timing results therefore hold whether the feedback weights learn or not.

Q3. "How robust is the proposed method to different initializations of the forward and feedback weights?"

We swept both (Appendix E.3, Table 4): forward weights at 0.25 to 4 times the paper's scale; feedback weights starting with equal entries (the paper's start), normally distributed or with random signs, at 0.5 to 2 times the paper's scale, learned or fixed; five seeds; the two-hidden-layer circles network and 7×7 MNIST, at zero delay and with a late label.

  • The feedback start matters little. At the paper's forward scale every feedback start and scale, learned or fixed, reaches 0.92–0.98 on the circles and 0.86–0.89 on MNIST at zero delay, and no run with learned feedback stalled.
  • The forward scale matters more, as in most networks. Very small forward weights slow learning: at a quarter of the paper's scale the circles network needs larger feedback weights to leave the majority class within the training time. Large forward weights are harder for learned feedback: at two to four times the paper's scale most of its runs stall in the solver, while fixed feedback learns at every scale. With the standard He initialisation the paper uses, learned feedback trains reliably from every feedback start.

I thank the authors for the additional experiments and clarifications. Understanding how inference and synaptic plasticity operate concurrently in neural networks is an interesting problem, and I appreciate the new insights provided by this work. I am therefore willing to raise my rating to 3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.