acceptodds
Under review as a conference paper at ICLR 2027

Meta-Learned Stateful Plasticity: Local Rules for Online Learning

Abstract

Biology motivates learning rules that modify the connection between neurons using only local information available at the synapse. While hand-designed local plasticity rules are interpretable, they are not generally optimized for task objectives. We present Meta-Learned Stateful Plasticity (MLSP), a framework for discovering local rules in online classification settings. The rules are parameterized by multilayer perceptrons (MLPs) that are shared across synapses and take in only locally available information such as pre- and postsynaptic activations, the current connection weight, and a broadcast reward signal, while also updating a writable neuron-local plasticity state. We apply MLSP to the hidden weight matrices of a feedforward network whose readout weights learn with the delta rule. The meta-trained rules outperform all tested classical rules, generalize to task families not used in meta-training, and are competitive with, though generally below, online backpropagation. Rules meta-trained with plasticity states reach higher average accuracy than rules without them, and we present evidence that the states track progress through the episode at the layer level. Finally, we fit polynomials of local variables to rules meta-trained without plasticity states. Networks trained with these polynomial rules retain much of the advantage over learning with the delta rule alone, but the fits do not reduce to any classical rule. This indicates that effective local learning rules extend beyond the hand-designed forms typically considered.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Local Learning Rules, Biologically Plausible Learning, Meta-Learning, Weight Transport Problem

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_P9PVOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

The paper proposes a meta-learned plasticity rule for a feedforward neural network classifier, parameterized by an MLP that is a function of the pre- and post-synaptic activity, the current synaptic weight, reward (whether the current step's class output is correct), and a per-neuron "plasticity state" that is updated at every step (but not used in the forward computation). The rule is meta-learned and evaluated on a suite of online learning tasks (permuted-MNIST, RHM, permuted-Fashion-MNIST, and Omniglot), demonstrating performance superior to classical local plasticity rules and in some cases comparable to online backprop.

Critical strengths
  1. The work towards mechanistic understanding of the rule through fitting it with an interpretable (polynomial) family, as well as perturbation of learning signals (plasticity state, reward), and the step-index control providing insight to the function of the plasticity state.

A minor note, although the paper was easy to follow without it, I appreciated the thorough glossary in the appendix.

Critical weaknesses
  1. Lacking comparable baselines with learned plasticity, e.g. meta-learning (rather than grid-search) the parameters of the classical rules, meta-learned polynomial fits (with or without feedback; see Questions 1-2), as well as including the version without plasticity states in Fig 3A-D.
  2. The complexity of the rule appears to be only partially justified. From the authors' analysis, the per-neuron plasticity state appears to contribute relatively little unique information beyond a timing signal, and even causes deterioration for learning beyond the trained episode length unless clamped/reset; and the reward is essentially unused.
Questions for authors
  1. What is the rationale for quadratic polynomial fit for the stateless rules, but cubic for stateful, and moreover using L1 for stateless but not for stateful? Even in the cubic expansion there are sufficiently few terms that make fitting straightforward, and would obviate the need for an ad-hoc appending of an Oja term to the quadratic approximation.
  2. Relatedly, why fit the polynomial to the learned rule instead of parameterizing the rule with a polynomial during meta-learning? Including the meta-learned polynomial as a baseline could strengthen the performance comparison in Fig 4.
  3. How much of a role does randomizing P per-episode play? Is this really necessary, or can the task network operate directly on inputs, or with a fixed P (and how would that affect generalization)?
  4. How important is it to train on the combination of pMNIST and RHM for generalization vs within-family performance?
3 comments
AuthorsOpenReview ↗

Response to reviewer P9PV: Weaknesses and questions 1 to 3

Thank you for the careful review and helpful suggestions.

Weakness 1. Thank you, we agree that these are useful comparisons. We meta-learned the learning rates (and additional parameters for BCM) of the classical rules.

We did this in two ways, i) initializing the hidden rates at a small value, 10−610^{-6}, and ii) initializing at the optimum of the grid search. Please see Appendix G.1 (Figure 19) for the performance of each rule.

Meta-learning the learning rates did not substantially improve baseline performance found by grid search.

For both initialization cases, we find that the comparisons agree with the grid-searched versions. One exception is feedback alignment on RHM, which performed worse when initialized at small hidden rates, but performed nearly as well when initialized at the grid search optimum.

We agree that rules without plasticity states should be added to Figure 3A–D. We have done this in the updated manuscript.

Weakness 2. We agree with both observations. It was an outcome we did not expect. The plasticity states appear to mostly keep time within the episode, the rule barely uses the reward, and the learned rule is largely unsupervised in the hidden layers. We have performed additional studies on the effect of the reward. In line with MLSP's relative reward insensitivity we find that a polynomial rule meta-learned without the reward performs, relative to the readout-only baseline, about as well as a polynomial rule meta-learned with the reward (Appendix G.3, Figure 24).

We agree that learning beyond the trained episode length is a real limitation of this work, as mentioned in the main paper Discussion. For future work, we would be interested to use much longer training windows or truncated BPTT to extend the learning window beyond the rapid online learning setting.

Question 1. Our initial motivation was to use the simplest interpretable fit that retained much of the rule's performance. In the case of the stateful rule, we found that a quadratic-plus-Oja and a cubic fit were insufficient, which is why we considered a modulated cubic fit (Appendix D.4).

We agree that for consistency the same cubic fit should be applied to both. We fitted a full cubic polynomial (all 20 terms of degree at most three in presynaptic activation, postsynaptic activation and weight, Oja's term among them) to each rule (i.e. random seed initialization) of both the stateful and stateless MLSP by least squares, with no penalty and no added term.

For rules without plasticity states, the cubic explains 73% and 76% of the held-out variance of the weight updates of W(1)W^{(1)} and W(2)W^{(2)} (mean of ten rules), against 68% and 70% for the quadratic-plus-Oja fit. We have added the details of this fit to Appendix G.2 (Figures 20 and 21).

Crucially for the conclusions of the paper, in neither case do we see the fitted polynomial reduce to any known classical rule.

Following questions from other reviewers, we investigate a task setting similar to Confavreux et al. (2020), where the task objective is extracting the first principal component. We find that the rule meta-trained with an MLP basis is close to Oja's rule when we fit a polynomial to it (Appendix G.10, Figure 32).

Question 2. This is a very good suggestion, and we have attempted to fulfill it within the rebuttal timeframe.

We meta-learned with a quadratic and a cubic polynomial basis and have included the results in Appendix G.3 (Figure 23).

After 100,000 meta-training episodes, both gain less over readout-only learning than MLSP using an MLP basis (with or without plasticity states) on all four task families.

Question 3. This is a good point. Because of the choice of task families, randomizing P is not necessary. If we were to meta-train with a task whose input distribution the rule could exploit, a fixed P may hurt generalization, but in this case it was additional randomization that we don’t expect to make much difference.

We test this by running additional seeds with rules meta-trained with one fixed P, and find they are as accurate as rules meta-trained with a new P every episode, whether tested with their own P or a new one. The results have been added to Appendix G.4 (Figure 25). As suspected, for the task families initially tested, redrawing P each episode did not make a difference (regardless of whether P is redrawn during evaluation). The same holds for the XOR task, which was added at the suggestion of another reviewer. Rules meta-trained for a network without P, in which W(1)W^{(1)} receives the 784 inputs directly, gain less over readout-only learning (Appendix G.4, Figure 26).

AuthorsOpenReview ↗

Response to reviewer P9PV: Question 4

Question 4. To answer this question, we meta-trained rules on pMNIST alone and on RHM alone. The results can be found in Appendix G.5 (Figure 27). The combination makes little difference within a task family, but does appear to matter for generalization across task families. For example, training only on RHM appears to prevent the rule from generalizing to Fashion-MNIST and Omniglot.

Thank you to the authors for the additional experiments and clarifications. I have raised my soundness score from 2 to 3.

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

This work uses meta-learning to construct new local rules for online learning: essentially it uses an MLP (described in Eq (2)) to parametrize an update rule for the weights in a neural network, one which depends on only local structures (i.e. the local activations and "plasticity states" which are extra scalar degrees of freedom which presumably are used to determine how much the neurons should change). This MLP is trained using a meta-learning outer loop. It is shown in simple tests that the local rule learned in this way outperforms most simple baselines (e.g. Hebb, Oja, etc.) and even comes close to online backpropagation.

Critical strengths

This is a very clean, simple and well-motivated construction. In my opinion it is executed very well: the update rule and philosophy are cleanly spelled out, tests are done against numerous baselines, and the utility of the various components (e.g. the local plasticity) is shown through ablations. I was somewhat surprised that it had not been done before, but as far as I can tell, after some investigation of the literature indeed it seems this exact niche is not populated.

Critical weaknesses

Though a clean and neat result, I'm not certain about the broader significance of this work, as philosophically similar results (though, as mentioned above though mostly involving the backwards propagation of errors) do already exist in the literature.

Questions for authors
  1. Regarding the overall weakness mentioned above: could the authors place this result in a broader context, and perhaps say why this is significant, what it might tell us about biological credit assignment, what they anticipate it being used for, etc.?
  2. Clarification request: why is the accuracy on pMNIST so low in Figure 3A? Even a fully linear model obtains 92%, and I would expect the backpropagation answer to approach that. Is this related to the evaluation protocol, i.e. does the training not allow the system enough time to approach good performance?
  3. Future work suggestion: as discussed, it is somewhat difficult to interpret the learned rule. Is it possible to study an extremely simple setup (say two neurons representing the XOR function or something similar) and then carefully dissect the resulting update rule to see if it has any universal properties? Can the authors explain whey their rule seems more difficult to interpret than that of Confraveux et. al?
  4. Clarification request: on line 72, it is said that the rule trained on pMNIST and and the RHM transfers to Fashion MNIST and OmniGlot, but as far as I can tell this is not explained in any great detail later. Could the authors explain to what extent the rule generalizes? For example, could one train a rule on an extremely simple few-neuron task and then use it on pMNIST, or something of that kind? How different are the rules learned for different tasks?
4 comments
AuthorsOpenReview ↗

Response to reviewer QiyS: Weakness and questions 1 to 3

Thank you for the assessment and the thoughtful suggestions.

Weakness. Thank you for raising this point. We answer it under Question 1.

Question 1. For a broader context, synaptic plasticity in the brain is local. A synapse has access to pre- and postsynaptic activity and to its own strength, possibly modulated by neuromodulators (broadly diffused chemical signals, perhaps conveying reward) and by an activity "local-field" believed to resemble the average neuronal activity; this last signal could be regulated by homeostatic plasticity. Since Hebb's postulate that neurons that fire together wire together, only a few other local learning rules have been proposed. Hebbian learning is intrinsically unstable. Oja's rule stabilizes it and, in the simple setting of a single linear output neuron, extracts the first principal component of its inputs (Oja, 1982). The BCM rule stabilizes Hebbian learning by a sliding threshold that allows the synapse to change between potentiation and depression. Three-factor rules, which also use a reward signal, have been studied at length as well. These rules are useful models of biology, and can explain various experimental observations and phenomena. They were not designed to solve tasks, however, and most are not able to solve non-trivial tasks such as classification to high accuracy out of the box, as we show in this work. It is of fundamental importance in neuroscience to understand which local interactions allow biological networks to learn effectively. We consider this work an early step in that direction.

As you point out, the closest meta-learned rules send errors or learned messages backward through the network. MLSP sends no errors backward, as each hidden connection uses only its two neurons’ activations and plasticity states, its own weight and one reward bit shared by all connections. There is quite a large diversity of different approaches to meta-learning local rules, but we believe this work occupies an unfilled niche.

For significance, we were surprised that a rule without neuron-specific error feedback comes close to online backpropagation on pMNIST (67.9% vs 71.1%), slightly exceeds it on Fashion-MNIST (63.0% vs 61.9%) and works on task families that were not used in meta-training.

For uses, we can speculate that it may be a source of hypotheses about which local signals biological plasticity needs (here mainly postsynaptic activation and the weight), and a first step to set up a framework for discovering useful rules in more biologically relevant architectures such as state-space models or nonlinear recurrent neuronal networks. For machine learning, we hope it may be considered a starting point for online learning without backpropagation.

Question 2. Yes, it is the evaluation protocol. Accuracy is measured on the 512th image, after learning from only the 511 images before it, each seen once, in a newly initialized task network. The 92% for a linear model typically uses all 60,000 training images, batched, and with many more training steps. Also, by narrowing the network to 64 units through several layers, the online learning performance depends more strongly on the plasticity of the weights connecting the hidden layers, which is what we are most interested in. But it is true to say that in a different setting a network with one hidden layer, or even a linear network, can learn the MNIST task quite well.

Question 3. Thanks for this interesting suggestion. We meta-trained three rules on XOR in a small version of our task network, with 2 inputs, hidden layers of 8, 4 and 4 neurons, and 48 plastic connections. Accuracy at step 512 is 81%, vs 69% for readout-only learning. Please see Appendix G.6 (Figure 28). We found that the learned rule is quite simple in form. A weight increases while its postsynaptic neuron is less active than a set level and decreases while it is more active, which keeps each neuron's average activity near that level. This holds for both plastic weight matrices and all three rules, and a rule of only this form reaches 80%.

With respect to the question of interpretability: In Confavreux et al. (2020), they rediscover rules such as Oja's by meta-learning the coefficients of a polynomial basis in the exact same setup in which Oja's rule was shown to extract the first PC of the data. We now replicated their setup, learning online, but instead of a polynomial basis we used an MLP. When we then fit a polynomial with an L1 penalty to the MLP, we find that it indeed recovers Oja's rule (Appendix G.10, Figure 32). So it really is that classical local rules are poorly suited to learning representations for rapid online classification, as our comparisons show.

AuthorsOpenReview ↗

Response to reviewer QiyS: Question 4

Question 4. On the point about generalization, since this is core to the paper, we agree it needs clarification. We have added the transfer accuracies to the introduction (Figure 3C,D), and further tests of generalization in Appendices G.5 and G.7.

Fashion-MNIST and Omniglot are task families that are never shown during meta-training, and so applying the final learned rule to these tasks is the strongest test of generalization. For pMNIST and RHM, we are relying on the extremely large number of possible permutations and hierarchies in the task family (more than a billion in each case) to all but guarantee that a task is never drawn twice from the same family. It is a weaker test of generalization, as these task families are used for meta-training.

We appreciate the suggestion of using the XOR task in the previous question and have used it again here to further test generalization of MLSP in the paper's standard architecture, where each half of the inputs represents one of the two input neurons of XOR. We find that meta-training on XOR does not generalize to the other task families. However, MLSP trained on pMNIST+RHM performs better on XOR than readout-only learning. The results can be found in Appendix G.5 (Figure 27) and Appendix G.7 (Figure 29). We have also included the other generalization combinations, meta-training on pMNIST alone and on RHM alone (Appendix G.5, Figure 27).

I am grateful to the authors for the detailed response to my comments and for implementing my suggestions on the XOR task and testing generalizability. I feel that all of my questions have been answered to my satisfaction, and further I think that the extra context added in the response to Q1 answers my primary concern about significance.

I think this is really interesting work and should be published. I am increasing my score to a 4 for strong accept.

I encourage the authors to further consider whether they can interpret why generalizability from the XOR task (e.g.) is so poor, I think this could lead to some interesting insights into how biologically plausible mechanisms might end up being different from task to task.

Thank you for the kind words and for this interesting suggestion, which we plan to look into if we are given the opportunity to provide a camera-ready version of the paper.

Reviewer_ZYDmOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents MLSP, a meta-learning approach for finding local learning rules. It uses MLPs to parametrize the rules as plasticity states associated with layers. These MLPs are trained after a batch of episodes, in each of which a different task from the same task family is learned. The paper motivates this work by describing existing work in local learning rules as being 1) hand-designed and 2) not performance-oriented. It shows that MLSP can learn across a task family, beat local learning baselines, and even have some meta-learned rules transfer to other task families. The paper conducts significant analysis of the learned patterns.

Critical strengths
  1. The paper is quite well-written - the language is strong and efficient. I marked off a point for clarity mainly for density, but this is pretty fixable - the writing style, structure, and overall approach to communication are especially good in the current pool of papers.
  2. The analysis sections are especially interesting. I appreciate the interpretability approaches used to investigate what matters for learning strong rules. I also think the experimental suite is well-designed.

The overall method also seems interesting and insightful to me, though I am less familiar with the state of the art in this field. So I would say that is a baseline strength and the main reason I recommend acceptance.

Critical weaknesses
  1. In some parts, like figure 1 and section 3.1, it's hard to build intuition for how the algorithm works. The content is laid out clearly but it's very detail-oriented and doesn't offer the mid-level idea - we go straight from high-level to low-level. Algorithm 1 does help build the intuition so it may be worth referencing it or placing it earlier.
  2. I am less familiar with the most current literature in this domain so this may be my lack of understanding but the fact that the meta-learning is nonlocal seems to go unexplained.
Questions for authors

I realize that the primary motivation for this work is to create local learning rules that are strong at learning the task, but I assume the restriction to locality is ultimately about biological plausibility. Does the meta-learning being nonlocal compromise this? If so, how? If not, why not?

1 comment
AuthorsOpenReview ↗

Response to reviewer ZYDm: Weaknesses and question

Thank you for the encouraging review and the helpful advice. We have tried to add some clarity to Section 3, in response to the suggested weaknesses.

Weakness 1. Thank you for this suggestion. We agree, and Section 3 now opens with a description of one episode step, which refers to Algorithm 1. We have made efforts to add intuition in the paper in two ways. First, by studying a simpler case. In Appendix G.10 (Figure 32) we use our method to learn the first principal component of the inputs, instead of performing online classification, and the cubic polynomial fitted to the rule we obtain is close to Oja's rule, as theory predicts. Second, we have added meta-learned polynomial rules as a comparison (Appendix G.3, Figures 22–24). We will continue to work towards improving clarity in the main text.

Weakness 2. Please see our response to the question below.

Question. Yes, you are right to point out that the MLSP method for finding a biologically plausible rule is not itself biologically plausible. In nature, useful local plasticity rules would have been discovered by evolution. However, the plausibility of the rule itself is not compromised by this, as within an episode the rule is local and each hidden connection uses only quantities available at that connection. A sufficiently flexible local rule, such as the ones that may underlie changes in synaptic strength in cortical circuits, would have taken a long evolution to find, but would generalize easily to unseen tasks. Finding such rules by an evolutionary search would be an interesting alternative for future work.

Reviewer_1Q8oOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates whether synaptic plasticity rules can be learned rather than manually designed. The authors propose Meta-Learned Stateful Plasticity (MLSP), which uses an MLP to learn how to update synaptic weights and neuron-local states from neural activity, connection weights, and broadcast reward signal. Experiments on permuted MNIST and the Random Hierarchy Model show that MLSP outperforms classical plasticity rules, and the rules can also transfer to Fashion-MNIST and Omniglot without additional meta-training. Furthermore, the authors try to improve the interpretability of the learned rules by approximating them with polynomial expressions, which retain much of the performance and do not reduce to classical plasticity rules.

Critical strengths
  1. Efforts toward interpretability. The polynomial fitting experiments provide a useful first step toward understanding the learned plasticity rules and identifying how they differ from classical plasticity rules.
  2. Careful experimental analysis. The paper includes ablation studies of local signals and plasticity states, as well as longer-horizon evaluations, providing useful insights into the behavior and limitations of the learned rules.
Critical weaknesses
  1. Insufficient comparison with task-optimized learning rules. The paper mainly compares MLSP with classical hand-designed plasticity rules, which are not optimized for the specific tasks considered. Since MLSP is extensively meta-trained, outperforming these rules is not particularly surprising. Without comparisons with simpler meta-learned alternatives, such as directly optimized polynomial rules, it is difficult to assess whether the additional complexity of MLSP provides a substantial methodological advantage.

  2. High computational cost and unclear scalability. Unlike simple hand-designed rules, MLSP applies an MLP to every plastic synapse after each sample, potentially introducing substantial computational overhead. The polynomial approximations reduce complexity but suffer noticeable performance degradation, while the original MLP lacks the interpretability of classical rules and has no demonstrated biological basis.

Questions for authors
  1. Previous work, such as Confavreux et al. (2020), has explored directly optimizing parameterized local plasticity rules, although under different task settings. Could the authors include a simpler meta-learned baseline under the same experimental protocol? In particular, directly optimizing the polynomial rules with temporal modulation, rather than only fitting them post hoc, would help establish whether the additional complexity of MLSP is necessary.

  2. How well does MLSP scale to larger and more complex networks? The current experiments are limited to a small feedforward network with two plastic hidden layers. Evaluations on larger architectures and more challenging tasks, together with an analysis of the associated computational and memory costs, would provide stronger evidence of the method's scalability and practical applicability.

  3. Could the authors provide additional evaluations on held-out samples to better establish the generalization ability of the learned rules?Although the authors demonstrate transfer to task families, the high sample repetition rates on RHM and Omniglot may inflate the reported accuracy and weaken the evidence for generalization to unseen inputs.

4 comments
AuthorsOpenReview ↗

Response to reviewer 1Q8o: Weaknesses and question 1

Thank you for the review. We address the concerns raised below.

Weakness 1. Thank you for this feedback. We agree that the lack of established task-optimized local rules makes it difficult to assess the advantage of our method. We also agree that a simpler meta-learned rule, such as a polynomial in place of the rule MLP, is a useful addition to our comparison with optimized known local rules (Appendix C.2, Figures 6–9).

We have now added this comparison. We meta-learned the coefficients of quadratic and cubic polynomial rules, used in place of the rule MLP and meta-trained in the same way as MLSP (Appendix G.3, Figures 22 and 23). Section 4.2 of the main text now refers to these results (line 367).

The simpler polynomial rules perform worse than MLSP with its rule MLP on all four task families.

Weakness 2. We agree that meta-learning by backpropagating through online learning is computationally expensive by its nature. This is a problem often faced when meta-learning in general. The advantage is that once a rule has been found, applying it is relatively inexpensive. The MLP is not applied to each weight in turn, but to all synapses of a layer at once, which take the place of the usual batch dimension of a forward pass. We address whether the rules found generalize to larger networks in the answers below.

We furthermore agree that two questions are fundamental: i) what a scaling law for meta-learning with MLSP might look like, for example how the accuracy of a fixed rule changes as the network grows or the number of episode steps increases, and ii) how the computational cost of meta-training scales with width and depth. Here are partial answers that we could provide within the rebuttal timeframe. Regarding i), the paper already shows that the same rule, applied without retraining at half and double width, still learns on every task family. Its accuracy rises with width on RHM and Omniglot, but not on pMNIST or Fashion-MNIST (Appendix F.2, Figure 18). In episodes twice as long, accuracy is largely kept if the plasticity states are reset or held fixed after step 512, but falls if they keep running (Appendix F.1, Figure 17). Regarding ii), we show in Appendix G.9 (Figure 31) how the time and memory of meta-training scale with width and depth. As this work is, to our knowledge, the first to meta-learn one shared local rule without neuron-specific feedback for online classification, we consider scalability as a direction for future work.

Question 1. We thank you for this excellent suggestion. We have included additional experiments (Appendix G.10, Figures 32 and 33) in which, instead of classification, the objective is to find the first principal component of the inputs, as in Confavreux et al. (2020). In Figure 32, we match the architecture used in their paper, a linear network with a single output neuron and no hidden layer, but learn online, one input per step. In this case, with an L1 penalty, the cubic polynomial fitted to the meta-learned rule recovers Oja's rule, showing that our method can recover a rule that was designed mathematically to achieve an objective.

From deep learning theory we know that the existence of a hidden layer can make a big difference to what is learned in the last layer. We therefore repeated the experiment in a network with one hidden layer (Figure 33). There the clean recovery of classical rules no longer holds. In the first layer, with an L1 penalty, Oja's rule terms are still the largest, whereas the second layer involves other terms and has no Hebbian term. Intriguingly, at the end of the episode these rules have a lower error than Oja's rule in the linear network. So by moving from a linear network to a network with one hidden layer, we immediately find rules that are harder to interpret in terms of known classical rules.

We would like to stress, however, that for the rapid online classification tasks we analyze, Oja's rule does not perform above readout-only learning in any task family. We consider our work to be complementary to Confavreux et al. (2020).

In response to the second part of the question (a point also raised by another reviewer), we have now also meta-learned the coefficients of a second and third order polynomial basis directly. We find that the polynomial rules gain less over readout-only learning than MLSP with an MLP basis on all four task families (Appendix G.3, Figure 23). We also meta-learned a cubic with temporal modulation, multiplied by a meta-learned decay in time. We run 20,000 meta-training episodes and analyze the gain over readout-only learning. We find MLSP gains more over readout-only learning than this cubic on pMNIST and RHM, the two task families used in meta-training (Appendix G.3, Figure 24).

AuthorsOpenReview ↗

Response to reviewer 1Q8o: Questions 2 and 3

Question 2. The rule is shared by all hidden connections, so it can be applied directly to larger networks. Of course there is no guarantee a priori that it would scale to arbitrary changes in size. Appendix F.2 studies this for networks with half and double the width. We do agree that understanding how our method depends on the network architecture, particularly for RNNs, would be interesting future work. Unfortunately, such a study will not be possible to complete within the rebuttal timeframe.

Question 3. Thank you for this suggestion. We added this evaluation. We show accuracy on samples held out not only from meta-training but also from the episode of online learning. MLSP still clearly exceeds readout-only learning on all four task families. The results can be found in Appendix G.8 (Figure 30). As you would expect, samples not seen during online learning are not classified as accurately as those that are repeated, but that is true for every rule, including online backpropagation (Figure 30B).

We would like to emphasize that this concern is with regard to seeing the same sample multiple times during online learning, which is true for all comparisons too, and is not relevant to whether a rule generalizes to unseen inputs, as the rule remains fixed during an episode of online learning. As you point out, seeing the same sample multiple times during an episode is more likely for Omniglot and RHM, given the small number of samples per task. However, what is important for generalization is that specific tasks are not seen during meta-learning. For Omniglot, no task is ever presented during meta-training, and each RHM hierarchy is also almost certainly new, as meta-training drew about 100,000 hierarchies, from more than a billion per configuration.

The authors provide extensive ablation studies and explanations. However, the current experiments and conclusions are insufficient to support claims of performance on more advanced, meaningful tasks.

Thank you for your reply and for noting the extensive ablation studies and explanations. In the absence of further comments or questions, we hope you agree that our responses to the two critical weaknesses and points raised in the questions were satisfactory (other than a remaining issue, below).

Regarding "However, the current experiments and conclusions are insufficient to support claims of performance on more advanced, meaningful tasks". We first want to make clear that we do not make such claims of performance on more advanced tasks in the manuscript. Second, it was not clear to us from the text of the review, and neither was a task suggested, that task complexity was the review's dominant concern - which we assume to be the case given your score remains at 1 (confidence 4). We would have run experiments earlier in the discussion period when there was time to do so if this was apparent to us.

More generally, in the literature on meta-learning local plasticity rules, relatively simple and often synthetic tasks are very common and it is difficult to find a meta-learned local rule that tackles more advanced classification tasks than the ones studied here (e.g. Confavreux et al., 2020; Sandler et al., 2021; Shervani-Tabar & Rosenbaum, 2023). As such, to our knowledge, the tasks considered are comparable to or more advanced than those studied previously in this area.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Revised paper

We thank all reviewers for their time and helpful comments. We have uploaded a revised version of the paper, with changes shown in blue. Additional experiments have been added to Appendix section G.

AuthorsOpenReview ↗

Summary of rebuttal revision

We have uploaded a new revision for the end of the reviewer-author discussion period. We would like to highlight some of the additional experiments now included in the submission, in response to questions raised in the reviews:

  1. Reviewers 1Q8o and P9PV highlighted the lack of a meta-learned polynomial baseline. We have included experiments showing that meta-learned polynomial rules do not perform as well as MLSP (Appendix G.3, Figures 22–24). Relatedly, Reviewer P9PV noted that the parameters of the baseline local rules were set by grid search, and suggested meta-learning these too. We include these experiments in Appendix G.1 (Figure 19) and find no substantial improvement over the grid-searched parameter settings.

  2. We have added experiments in which we apply MLSP to the setting of previous work, Confavreux et al. (2020), and learn to extract the first principal component (Appendix G.10, Figure 32). Here, the rule found by MLSP is interpretable, and corresponds to Oja's rule, in line with previous work. This was in direct response to Reviewer QiyS Question 3, and we think it is a meaningful positive control of broad relevance to the reviewers.

We also appreciate the discussions we have had with the reviewers on a more individual basis. We thank the reviewers for their time and the set of constructive reviews. We believe the manuscript is strengthened as a result.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.